The Cheat Sheet

(438) New York Times Covers AI and Academic Cheating

Derek Newton — Thu, 18 Jun 2026 23:05:08 GMT

Subscribe below to join 5,377 (+13) other smart people who get “The Cheat Sheet.” New Issues nearly every Tuesday and Thursday.

The Cheat Sheet is free. Although, patronage through paid subscriptions is what makes this newsletter possible. Individual subscriptions start at $8 a month ($80 annual), and institutional or corporate subscriptions are $250 a year, suggested.

The Cheat Sheet is verified human-written:

The NYT on Cheating and AI

When I see coverage of academic fraud, such as this piece in the New York Times, in the world’s top media properties, I have mixed feelings.

I am usually really quite happy for the visibility. Lack of awareness and its indirect pressure to act have made it far too easy for far too many people to do nothing about academic cheating. So, when headlines pop in major places, I applaud.

At the same time, I think I know too much. The errors, though usually minor, really upset me. I feel as though, if we’re going to do this, we should do it right. Especially at places such as the NYT.

On balance, this is a solid article. It does not miss much — maybe just two things I’ll point to in a second. I guess the second is up. Here are my two issues, so we can get to what’s really good about this coverage.

Misses

Unfortunately, the biggest issue is the headline, which is all that most people will read. My issue is that it’s wrong. Here it is:

Student Cheating Is Becoming Impossible to Detect in an A.I. Era

Not true.

There are companies, which are really the welcome focus of the article, that say that they make cheating impossible to spot. But they’re selling product. And I do not think there is any evidence to support such a statement as true — that “cheating is becoming impossible to detect.” The article does not provide any.

This may be a long-standing and frustrating hole in journalism wherein the article’s author, the journalist, does not write the headline. This happens. When it does, a rushed editor or junior-level digital manager can write it, leading to errors.

Still, getting the literal headline factually wrong — or, at best, unsupported — is a cringe.

My other quibble is that the story references a survey showing that:

About two-thirds of American students are using A.I. regularly for schoolwork, according to recent surveys. While only a small slice — about 9 percent — admitted to outright cheating in one large study, much A.I. use lies in an ethical gray area.

Obviously, there is zero chance that the 9% figure is accurate. And I don’t understand why it’s so difficult for writers in this area to point out, factually and accurately, that self-reported survey data about cheating is wildly low. People don’t admit to cheating. They don’t think that what they’re doing is cheating. Standing alone, any such number is misleading — at best.

I also hate that somehow using AI is now “an ethical gray area.” It is not. This is a hill on which I will die.

If you are using AI to deceive, to take credit for work that is not done by you, in violation of rules you are expected to follow, nothing whatsoever is ethical about that. I am genuinely baffled as to why anyone thinks otherwise and, likewise, do not understand why this idea is so frequently, so uncritically repeated.

Much Love

Anyway, the piece is still very good and important. And, aside from the headline, my quibbles are just that. In the very next paragraph after the 9% nonsense, is this:

A recent College Board survey of professors found three-quarters reported their students were using A.I. to write, and over 90 percent of respondents were concerned about plagiarism and dishonesty. Many institutions have seen a sharp increase in student disciplinary cases for academic misconduct, much of it related to the use of A.I.

Three-quarters.

Were you in Vegas, money in your pocket right now, which number would you be betting on? That cheating is closer to the 9% that students say it is? Or closer to the 75% that teachers report?

I know, the 75% isn’t identified as “cheating.” But, in your heart, where are your chips? On nine? Or 75%?

But what really sets this article apart, and what I love most about it, is that it exposes Grammarly as a cheating provider — my words, not theirs. At least not directly. And that the article goes deep into the companies overtly pitching, selling cheating services, by which I mean they are telling students that — wink, wink — you can use our services to get a good grade without doing any work whatsoever. There’s outstanding video of these throughout the article.

I also love that this article is among the first pieces I’ve seen to go into auto-typers, the software that can mimic human typing, creating a fake record of work, deceiving tracking software such as Google Docs and the educators who have come to rely on these records.

Bits and Bites

If you read The Cheat Sheet, not much in the NYT piece is really news. We’ve written about auto-typers, Grammarly, and brazen marketing before. All the time, in fact.

Still, I encourage you to read the reporting. It’s good. I’ll share a few snippets that I think are especially so. Here’s one. Running through how companies are selling cheating, The Times writes:

Even established ed-tech companies are marketing with a wink and a nod.

They then show a video clip from Grammarly. I’m here for every second of it.

There’s also this:

In some cases, the very same companies selling detection tools are also making apps that allow students to cheat, including by writing papers for them or rephrasing text written by others. The apps promise to help them avoid accusations of misconduct by scanning their work before they submit it, allowing them to rewrite passages identified as A.I. Even honest students are often willing to fork over $10 to $20 per month for premium tools, since A.I. detectors sometimes flag legitimate work.

I mean sure, if you use “A.I. detectors” as one thing, they do sometimes flag legitimate work. Especially when those same “detectors” are selling the “fix” for what they “find.” Cough, cough. Just amazing there is something that needs fixing so always.

But the next graph is a quote from Grammarly — to mention one company that has an AI detector while also selling the tools they claim will let you skip all that detection stuff. The very same company. Imagine that.

Of course, Grammarly says there’s no point to AI detection and:

urged educators to accept that most future writing would be produced in a partnership between artificial intelligence and human discernment.

Just accept it. Stop fighting.

Here’s another clip, close to the heart of the article:

[there is a] fiercely competitive market of legacy ed-tech purveyors and tiny start-ups, all using social media to tell young people that their academic lives could be easier — much easier — if they embrace A.I.

True.

It continues with:

Some start-ups explicitly teach students how to cheat.
Meanwhile, established companies often urge students to use their tools responsibly as aids for studying, research, brainstorming, outlining and revision. But many of them are simultaneously producing technology that can easily be used to plagiarize and cheat. They put out tongue-in-cheek ads alluding to their ability to help students get away with something.

Also true. Cheating is big business. Even the big companies think they’re being cute telling people to be responsible, while waving them through the cheating lines. For money, of course.

The coverage of autotypers is great too. There are many examples. Here’s one:

A TikTok video about another app, Typeflo, told students that they could relax, watch YouTube and eat a sandwich while their essays were produced for them.

Yup.

Another TikTok account, udoka_comet, features over a dozen videos of a young woman discussing Comet, an A.I.-powered web browser from the company Perplexity. In one video, the woman says she doesn’t feel like writing a five-page high school lab report. She shows how Comet can do all the work for her, calling it “magic.”

To their credit:

[a spokesperson for Perplexity] said the company had cut ties with an advertising agency that had “taken liberties” to increase online engagement, and that Perplexity had reminded social media partners to focus on “appropriate, responsible uses of Comet.” The browser can help with tasks educators might approve of, like formatting citations and creating study guides. But it can also complete assignments from start to finish.

Half credit.

Grammarly. Again.

From the piece:

Some professors are increasingly concerned about Grammarly, an app that has existed for 17 years as a sort of muscular spell-check. It now offers an “authorship” tool that helps professors screen for A.I. misconduct, by analyzing a document’s version history.
At the same time, the app allows students to generate writing from scratch, humanize text, and scan and replace phrases that could set off A.I. detectors.
Grammarly also provides a paraphraser that instantly rewrites any published text a student copies and pastes into a browser tab, which could be considered a form of plagiarism.
Grammarly advises students to use text-generation features “responsibly,” by citing each instance where A.I. was used in a paper. But the company also puts out ads that suggest students can use the app to pass off A.I.-produced writing as their own: “Detect A.I. text — it’s 2026, after all,” says one TikTok post. “Spot A.I. phrasing and choose edits that feel true to you.”

Yup. Grammarly sells cheating tools. Full-on, cheating. Been saying it. Going to keep on saying it.

Grammarly’s brain rot response:

Like other A.I. executives, Ms. Maxwell, the head of education for Superhuman, which makes Grammarly, said cheating has always existed but represents only a small segment — she estimated 10 percent — of student A.I. use.

People have always made war. We just sell guns. What can you do?

It is neither an ethical nor logical response.

Grammarly goes on:

“I can’t solve the human behavior issue that is cheating or pushing the easy button,” she said. “It is out of our realm.” [Maxwell said.]

Here’s a crazy idea. Maybe don’t make the easy button. Maybe don’t sell the easy button. Maybe don’t advertise the easy button.

One more:

George Cusack, director of A.I. academic initiatives at Carleton College, noted that Grammarly is sold to students as a benign helper when, in fact, “it’s a suite of tools that will do everything for you. It’s kind of shocking.”

You go, George.

GPTZero

I’ll end here, since this was news to me and therefore may be to you as well. From the coverage:

Some A.I. companies pitch themselves as protectors of academic integrity. One of those is GPTZero, which was born in 2023 as a Princeton senior thesis, and claims to be 99 percent effective in detecting A.I. content, including some use of humanizers and autotypers.
At first, the company marketed itself mostly to schools. But more recently, it has flooded TikTok with videos from purported educators, in which they explain to students how GPTZero will be used to reveal cheating and get them in trouble. The goal is for students themselves to use the app.
One social media user known as studyingwithjake describes himself as a graduate teaching assistant who helps students understand how professors use A.I. detectors.
“I want to show you what this professor’s been hiding from students,” he says, as he walks viewers through the interface of GPTZero’s browser extension. The tool analyzes a document’s version history, detects A.I. and can provide writing feedback. Once downloaded, however, users will find that it can also generate a full academic paper in mere moments, complete with quotes and citations.
“If you grade your paper this way before you submit it, you’ll probably get a good grade on your paper,” the influencer says.

Continuing:

The man in the video is actually an Arizona-based marketer named Jake Austin Sivilla, who wrote on LinkedIn that he created a fictional persona in order to win millions of video views for his client, GPTZero.
Mr. Sivilla declined an interview request, and his LinkedIn post was deleted after The Times began inquiring about the videos.

GPTZero responded:

Edward Tian, GPTZero’s co-founder and chief executive, said the company no longer worked with Mr. Sivilla and was moving toward working only with social media creators who are authentic educators or students. He also said there had been internal debate at GPTZero as to whether to allow the app to write school assignments from scratch, and that such capability might be limited in the future.
“Our mission is to preserve human quality and critical thinking in an age of A.I.,” he said.

Your mission is to preserve human quality but you’re having an “internal debate” about whether to let your tool be used to avoid that mission, explicitly. Gross.

Fine, One More

I’ll end where the NYT does, with one last quote from Grammarly:

Withholding A.I. entirely, [Maxwell] argued, was akin to educational malpractice, since students will be expected to use A.I. in the workplace.
“We’re doing a huge pedagogical upheaval in education,” she said, calling it a “burn it down moment. We’re just in the early stages.”

Grammarly and Maxwell are clear that they don’t believe there’s any point in resisting AI in education. Convenient that’s what they sell then, I guess.

Still, “burn it down” is a new low. Even for Grammarly, which is wink-and-nod peddling all the kerosene they can. They are not your friend. They are not an education company. They are a “burn it down” company.

Outstanding that the NYT has highlighted it.

Subscribe now

LA Times Runs Cheating Story Too — But Crazy

A few days before the outstanding coverage above from the New York Times, the LA Times ran a story on academic cheating too (subscription required).

Like the NYT piece, I am glad when this topic gets mainstream, high-level coverage. But also like the NYT offering, the headline is wrong:

Inside college AI cheating wars: extreme surveillance, false accusations, jarring confusion

Unlike the NYT example, I don’t mean it’s factually wrong. I mean it’s not journalism. For example, “extreme” is an adjective. It’s opinion. One person’s extreme is a second person’s quite reasonable. I probably am that second person.

To use a word like extreme, especially in a headline, you need someone — preferably an expert — to say it. The word appears in the article twice, both times attributed to no one whatsoever. It is the writer’s personal, uninformed opinion.

Let’s just say the piece has a view. As such, it’s advocacy. Not journalism. And I am not motivated enough to hack it apart, although it deserves several good whacks.

Here’s one. In the very first paragraph, the article says:

false accusations against students are increasing

No one in the article, no source in the piece, makes this point. It’s invented.

Back to the “extreme,” we get:

At UCLA, students in a recent sociology class said they were told in an email to “procure a mirror large enough to fully reflect your entire desk-area work space,” and turn on their laptop camera so the professor could watch them during an online test. In another course, students said they had to take their oral video exam with their arms crossed in front of them or behind their heads so they couldn’t type into AI platforms.

Is that extreme? Maybe. Although it’s not new. During pandemic remote assessments, many professors required two cameras on the desk and workspace — one front, one from the rear. Or mirrors. I saw that too.

Also, whatever, but not allowing students to type during an oral exam won’t stop them from using AI. Like 100 insidious companies sell services to feed you AI-powered answers on screen while you do nothing.

Here’s another easy chop:

The fallout has hit hardest upon students who said they did not use AI and were accused because of a professor’s suspicion or detection software that researchers have found can produce false-positives, especially on writing by non-native English speakers.

Saying the fallout — from a rise in AI use and related cheating activity, I have to assume — has “hit hardest on students who said they did not use AI,” is opinion. Again. No one in the article says this, let alone an expert. A contrary view is very easy to articulate. Plus, note please that it’s students “who said” they did not cheat.

Moreover, that’s the damn Stanford study again. No one has read it. It’s terrible (see Issue 216).

The article also trips over itself to show how students, scared and confused, are taking defensive measures such as using Google Docs and dumbing down their work. One does not work, the other is a waste of time.

I do like that the LA Times interviews an actual expert. Two or three, in fact. Here’s one:

Tricia Bertram Gallant, who directs academic integrity at UC San Diego, urged caution when describing the rapidly changing cheating environment, saying it reflects how poorly prepared colleges are to deal with AI use in classrooms.
“I would argue against the verbiage of ‘falsely accused,’ ” said Gallant,

I would argue against that verbiage too.

Gallant continues:

For Gallant of UC San Diego, the bigger problem is not widespread false accusations, but a system struggling to catch up with new technology.
[She said] “I have met with hundreds of students who have faced allegations of misusing AI and they don’t have that attitude of ‘Oh my god, I can’t believe I’ve been falsely accused.’

The expert literally says the problem is not “widespread false accusations.” The LA Times uses it anyway. In the headline.

Two more things, then I have to go. One:

At College of the Canyons in Santa Clarita, Adam Kaiserman said he reported a student last semester after his AI scanner spotted nine instances of possible cheating. He said the student’s work contained factual errors about the assigned reading, lacked quotations and was generic. But one of the cases collapsed after the campus investigator’s advanced AI detection tool disagreed with the professor’s AI scanner.

Unlike most of this story, this is a real problem. No matter what school policy is, teachers are using AI detectors. Problem is, most of the ones teachers use — the free ones — are garbage. Two different detectors can, and will, return different scores.

Rather than train teachers and give them access to good tools where everyone can agree on the flags and scores, schools have tried to shut the door, leaving teachers to freelance. It’s a terrible solution, if it’s a solution at all.

Finally, there’s this quote from Gallant, which should have been the story, but wasn’t:

“This fantasy that we can grant degrees based solely on unsupervised, unobserved work has finally, maybe, come to an end,” Gallant said.

I don’t know — that feels important. The LA Times went another way.

Subscribe now

(437) Two Research Briefs - AI Use in Math Assessments, AI Detection Accuracy

Derek Newton — Thu, 11 Jun 2026 17:57:27 GMT

Subscribe below to join 5,364 (+80) other smart people who get “The Cheat Sheet.” New Issues nearly every Tuesday and Thursday.

The Cheat Sheet is verified human-written:

Research Briefs — AI Use and Learning, and AI Detection Accuracy

Two new studies touch on academic integrity and are worth a moment or two of your time. One is important. The other, I am disheartened to report, is junk. Worse, it’s actually a disservice to intelligent, productive conversation.

I’m only going to skim them here, for speed and ease. The important one, I’m doing a longer piece for a mainstream outlet and plan to return to it in more detail here shortly. The other I may — or may not — return to. It depends on how angry I stay.

The Good One

A new study shows that AI use by students is showing up in key education indices including time on task and information retention. It also illustrates, once again, that assessment proctoring or similar anti-cheating interventions work.

The paper is “Faster Completion, Less Learning: Generative AI Reduced Study Time on Math Problems and the Knowledge They Build,” from authors Sina Rismanchian, of University of California, Irvine, and a team of four from McGraw Hill, the prominent education publisher and digital learning company. Those four are: Hasan Uzun, Jeffrey Matayoshi, Eric Cosyn, and Eyad Kurd-Misto.

The research team looked at millions of data of student engagement on ALEKS, “a widely deployed adaptive mathematics learning and assessment platform serving more than four million students annually.” Significantly, the data spanned more than a decade, giving insight into pre-AI activity, as well as post. It also crosses learning levels, from elementary school to college.

The results center around one key metric — time on task. That’s how long a given student spends from getting the question to giving an answer. It’s an important indication of effort and rigor, investment in the thinking, reasoning, and learning process.

From there, the paper breaks out two kinds of math problems.

They look at word-based questions which the researchers theorize are easy for students to flip over to AI. A simple copy-and-paste, cell camera image, or nested browser plug-in will spit out the answer pretty instantly. Low friction, low time on task, low effort.

They look as well at graphic or image-related math questions which they believe are harder for the AI bots to tackle. Not impossible. Just more of a trick or two is needed to get AI to skip the work for you.

Therefore, by looking at time on task for AI-susceptible questions, and those that are less so, across the pre-AI era and after, we’ve got a very interesting data set. And powerful results. At least I think so.

Findings

Before we all entered the land of GPT, time on task on the observed math assessments was steady, very consistent. There was no significant time/effort difference in which kinds of problems students were asked to solve. Not for years.

But after AI arrived and became everywhere all at once, easy to find, use, and hide, things changed. Time on task for the word problems began to drop. Steadily. Declining each quarter, one after the other, leading the research team to infer that these questions were, in heavy practice, being outsourced to AI during the assessments. AI was, after all, the only change.

From their paper (with some math symbols and equations removed):

[In the post-AI era] For college-level topics, learning time for word problems declines by 2.80% per quarter relative to graph problems, accumulating to a 26.9% cumulative reduction over eleven post-ChatGPT quarters. High school topics show a larger effect of 3.35% per quarter, corresponding to a 31.3% cumulative decline. Middle school courses show a smaller but statistically significant effect of 0.86% per quarter, accumulating to 9.0%. Grade 5 courses show no detectable change.

To summarize the decline in time on task since the wide release of generative AI:

College-level: down 27% overall
High school: 31.3% decline.
Middle school: drop of 9%.

The good news, I guess, is that, there’s been no observed change at fifth grade.

Proctoring and Prevention

Where proctoring is defined as both passive and active observation of the student during the assessment, the team found that proctoring and prevention work to limit use of AI. From the paper, math removed again:

In non-proctored assessments, response times for AI-susceptible topics decline by 1.11% per quarter relative to graph-based topics after ChatGPT’s release, accumulating to an 11.6% cumulative reduction. In proctored assessments, the estimated post-ChatGPT slope and cumulative effect are effectively zero and statistically indistinguishable from it. Event-study estimates confirm these results throughout the observation period, with parallel pre-trends in both conditions and a separation in the non-proctored setting beginning at ChatGPT’s release.

And:

The behavioral evidence establishes that students complete AI-susceptible problems faster in unmonitored settings in the age of generative AI, and that this behavioral change is absent under proctoring.

When students know they are being observed, when the risk is above zero, the cheating declines. Shocker.

Learning Loss

I hope you’re sitting down because, by adding questions to the assessments in proctored settings, where access to AI was reduced, the paper was able to measure learning loss, as evidenced by lack of retention, on previous questions that were susceptible to AI shortcuts.

From the paper, math removed again:

we estimate a 2.61% per-quarter decline in the log-odds of correctly answering AI-susceptible items relative to AI-resistant items in proctored conditions, accumulating to a 25.0% cumulative reduction in the odds of correct response over eleven post-ChatGPT quarters

This “retention decline”:

is the downstream consequence of the earlier learning phase, during which students appear to have relied on AI assistance to complete AI-susceptible problems without developing the durable understanding those problems were designed to build. This temporal ordering — AI use during unproctored learning, retention failure in subsequent proctored assessment — supports a causal interpretation of our findings.

A 25% cumulative reduction in learning since AI.

More on this really strong paper to come.

The Awful One

Another paper is out, once again making bold and unsupported claims about the accuracy of AI detection systems.

The title is:

AI detecting AI in academic writing: Why most AI detector findings are false

It ought to be retracted. And buried. Out of shame. But as sure as I draw breath, it not only won’t be retracted, it will be repeated, and repeated, and repeated. In nearly every case, by those who have not read it but are nonetheless invested in, and clinging to, the disproven idea that AI detection does not work.

The paper is by Panagiotis Tsigaris, of Thompson Rivers University in Canada, and Jaime A. Teixeira da Silva who is identified as “independent researcher” in Japan.

I’m going to try to avoid ripping this paper to shreds. For now. But here are three things to know about it.

One, the authors literally say that “most” detector findings are false. That’s a majority — at least 50% + 1. That’s wild beyond compare.

Especially because, two, they did not test any AI detection systems. None. Not one.

Three, the entire paper is based, as far as I can tell, on the single statement from OpenAI — the makers of ChatGPT — that their own detector was not accurate. That statement, by the way, was made in January 2023. That’s three and a half years ago — an eon in AI technology.

And it was never true in the first place. OpenAI’s detector may have been garbage, but the company lied and said all detection was garbage (see Issue 241).

Insanely, the writers of this paper take OpenAI’s ancient claims as true, never even considering the appalling business-serving case that OpenAI was making — that having good AI detection hurts their business of selling AI.

Three years ago, OpenAI said their own classifier was, from the new paper, quoting OpenAI:

“… its latest tool incorrectly labelled human-written text as AI-written 9 % of the time, and only correctly identified 26 % of AI-written texts”

That’s the rock on which this paper rests. That’s it.

From there, the authors assume that AI-edited text is harder to catch. They accept the claims that non-native speakers are more likely to have authentic writing flagged, though this is probably not true. They insist, without evidence, that AI technology is improving while AI detection is lagging. And, from that, starting with the 26% error rate they accepted from OpenAI — based only on that — they say most AI detection is wrong.

That’s not research. That’s PR. And wrong. Again — they tested nothing. Noth. Ing. But now there’s a paper saying that, “most AI detector findings are false.” Full sarcasm font — come on.

Maybe I’ll get back to this one. I’m not sure. I just don’t know how much patience I have.

Subscribe now

Dozens of Students Caught Using AI to Cheat English Exams in Australia

Nothing about this is news, I don’t think. It’s quite standard stuff. But it made the news, which is good. So, I am passing it along.

According to local coverage in Australia, about three dozen students at Melbourne's Mazenod College — which may actually be a prep school — were found to have used AI to cheat an English assessment. From the article:

Mulgrave's Mazenod College said it had recently discovered that 35 of its Year 12 students had used AI tools in an oral English exam, but would not confirm the exact number.

The school’s principal was quoted:

"While the use of AI tools is a growing challenge within all schools, they have no place in assessments and examinations, where every student must be able to demonstrate their own knowledge, independently and fairly."

Agreed. And not complicated.

Good for the school for not ducking this. Also good for them drawing a line and, it seems, holding students accountable.

Subscribe now

EdWeek: Young Students Are Turning to AI for Assignment Help Instead of Adults

EdWeek has a notable article out this week as well.

The short version is that:

Nearly a quarter of 9- to 17-year-olds say they would turn to a chatbot for help with schoolwork or homework before seeking guidance from a trusted adult such as a teacher, counselor, or parent.

That’s based on:

a new report from Common Sense Media, a nonprofit that researches and advocates for healthy tech use among youth.

I did not read the foundational report. Maybe I can get to it. But I left the link in, should you have interest.

Anyway, assuming the information was accurately reported and that the report is worthwhile, I’m surprised it’s just “nearly a quarter.” Then again, that’s probably those that admit to going to AI first. Even at nine, people know what the right, proper, expected answer is.

Euphemism alert — EdWeek says that:

That shift signals a growing reliance on AI as a first-stop resource for problem-solving and decision-making among young people.

Problem-solving. Sure. Fat chance.

EdWeek also says, based on the report:

The vast majority of 9- to 17-year-olds use AI. And of those adolescents, 85% have used AI for schoolwork or homework

Eighty-five percent. Problem solving. Go on.

Later, EdWeek describes AI as:

an educational tool

And it quotes a student who describes it as:

an indispensable educational resource

Uncritical hype much?

The article does mention accuracy concerns of AI-created answers, and privacy challenges of using AI, it does not have one word about cheating — about just getting answers with no work of learning. Because it’s a resource and tool for problem solving, and students are totally using it that way. Wink, wink.

Given what the paper up top shared about learning loss related to AI outsourcing, not even mentioning appropriate uses of AI is a big miss.

Subscribe now

(436) Academic Integrity Leader Uses AI to Advise Students "Don’t Cut Corners. Don’t Outsource Your Thinking"

Derek Newton — Thu, 04 Jun 2026 20:06:09 GMT

Subscribe below to join 5,284 (+2) other smart people who get “The Cheat Sheet.” New Issues nearly every Tuesday and Thursday.

The Cheat Sheet is verified human-written:

“Don’t Cut Corners. Don’t Outsource Your Thinking, However Tempting That May Be.”

Sometimes, life is stranger than fiction — too bizarre, too ironic to believe. This is one of those times.

The headline is that someone wrote an article for The Sydney Morning Herald (SMH) in Australia. The article included a suggestion that young people skip college because degrees were increasingly worthless. This triggered an article in response from Dr. Cath Ellis, Pro Vice Chancellor in Quality and Integrity at Western Sydney University. Ellis’s piece advocated for college, advising young people to, among other things:

“Don’t cut corners. Don’t outsource your thinking, however tempting that may be. If the system is as fragile as some claim, then genuine effort will not be hidden. It will stand out,” she wrote.

She advised students to, “do the work.”

Proving that there is no ceiling on irony, Ellis’s article was written with — or by — AI. After publication, the Sydney Morning Herald retracted it, pointing out that AI-generated articles violate their publication policies, especially when AI use is undisclosed, as was true here.

The Guardian and the SMH covered the retraction.

Making bad worse, Ellis is a legitimate leader, researcher, and advocate in academic integrity, especially on contract cheating. Her work has set the table for understanding and addressing the challenge for years. I’ve interviewed Ellis a few times. Outsourcing work, taking credit for work without work-related effort, is a topic she understands better than anyone.

The entire thing is bizarre.

But there’s more.

Ellis said she’d trained an AI bot to write in her style, uploading some 40,000 words of her own work. She then used the bot to turn her ideas and notes into full writing.

It’s just my opinion, but I see that is a distinction without any difference. If her advice to others in dealing with AI is to not cut corners, her path cut so many it’s now a circle. Ellis disagrees, telling the SMH:

But Ellis denied that her piece was written “by” AI. “It was written with AI, and there’s a really big difference there,” she said.

Sure.

Now, imagine this quote, from Ellis, coming from a student in an AI-related misconduct hearing instead:

“I really do feel that it’s allowed me to focus more of my time and energy on what really matters, which is the ideas, the thinking … rather than spending a lot of my time writing sentences from scratch.”

Do the work, amiright?

Many people, myself included, think that writing is thinking. If you’re not “spending a lot of time writing,” you are probably not spending enough time thinking.

And here is where the problem gets still worse. In a move they are sure to regret, Ellis’s school, Western Sydney University, defended her use of AI to “write” the article.

The spokesperson said the use of an LLM to draw on her own expertise and experience “demonstrates a sophisticated and appropriate use” of generative AI.

And:

“The University believes the AI use in this case was appropriate.”

Sophisticated and appropriate, right from the school itself. Good luck in those integrity hearings when students say they used AI to avoid spending a lot of their time writing sentences from scratch.

I am sure that 95% of today’s college students can train their own AI models with previous writing, then use that to skip the actual writing. What then? Remember when self-plagiarism used to be a thing? I do. Tell me how this is different.

Not rhetorical.

That cracking sound you may hear right now is the conferred value of a degree from Western Sydney University. If using AI to avoid writing sentences while submitting undisclosed AI as your own work is “sophisticated and appropriate,” what — aside from writing good prompts — is the school saying its graduates can actually do?

That is not rhetorical either.

By the way, even though the school was cool with Ellis’s AI use, as mentioned, the SMH wasn’t amused:

Herald editor Jordan Baker said the masthead had not been informed of the use of AI in the compilation of the article by either the author or WSU, and it did not meet editorial standards.

And:

“If our editors believe an author’s work is not entirely their own, it won’t be published.”

Standards.

Interesting.

If colleges and universities think that preparing students with “sophisticated and appropriate” uses of AI is what will help them get, and keep, jobs, they are seeing rainbows. Book publishers, media outlets, nearly all academic writing, consulting firms, financial service firms, investment houses, even AI companies themselves ban using AI in many circumstances, if not across the board. Every day, another company is embarrassed by its employees using AI to pass off cheap and easy as value added work.

Not everyone wants AI, in other words, so doubling-down on teaching it seems like it may well be a major miss. But I’m off topic.

Two more things.

One is simple. Training your own AI to write like you will not fool people. Good AI detectors catch that. Like, all the time. It did here, in addition to suspicion from others.

Two, as hard as all of this is to process, I further struggle to believe this, reported to have been from Ellis:

Ellis said these [AI guidelines at SMH] had never been conveyed to her. “If I’d known, I absolutely never would have sent it in,” she said.

She did not read the syllabus. I mean, the SMH AI policy is on their website.

In addition to advising students to read the rules and know the policies, they’re also encouraged to ask questions about those policies. Usually, nearly everywhere, not knowing the rules is no excuse.

Again, I can already hear students at Western Sydney and elsewhere saying, that they did not know that AI was banned and, had they known, they for sure would never have submitted AI-created work. For sure.

This is a debacle. An embarrassing, hypocritical debacle. So much so that it would not even pass for fiction.

Subscribe now

Inside Higher Ed, Again

It’s become an embarrassment how bad Inside Higher Ed is on the subject of academic integrity and cheating.

It seems the publication never misses any chance to dismiss cheating as a problem or to convince readers that the entire enterprise is hopeless. The dynamic is further complicated by the fact that IHE has taken ad money from cheating companies. More than once. But whatever the reasons, it’s beyond a pattern. At this point, it’s simply expected (see Issue 167).

And, so here we are again — IHE has blessed us all with another essay about academic fraud and cheating, wildly inaccurate and wrong without any fact-checking, oversight, or editorial care. Surprise, surprise.

This most recent one is by Kim Manturuk, “the executive director of the Center for Excellence in Teaching, Learning and Online Education at Georgia State University.”

That’s a shame because Georgia State has done some impressive things related to integrity. I consider them a good model in many ways. But this — this is something else entirely.

Her piece starts well enough:

If you work in faculty development, you have probably heard the same concern on a loop for the past year: All my students are cheating using AI. At Geogia [sic] State University, our campus teaching and learning center gets more requests for workshops on how to prevent digital dishonesty than any other topic.

That IHE missed the spelling of Georgia in the name of the author’s school says how much effort they’re putting into their editorial gatekeeping these days.

Anyway, yes, AI cheating is a concern. Rightly so.

But then we get the first tell:

We have all read the stories about professors reverting to blue books or opting for early retirement to avoid the perceived flood of machine-generated text.

Perceived flood? Really? That’s not reality? If you want to minimize something, debate how people see it, not what it is.

Turns out, the tell is right on the mark, as our author continues:

higher education has largely retreated into two defensive postures: surveillance or supplication.

With “surveillance,” we are off to never-never land:

The surveillance strategy relies on detection, an arms race we have already lost. AI-detection tools are biased, easily circumvented and prone to false positives. To test this, I fed the first chapter of my dissertation (written in 2006) into a popular AI detector. It flagged my work as 39 percent AI generated. We cannot police our way out of this when our radar is broken.

Wrong. Simply wrong.

The evidence that AI detection tools are biased is scant, or absent. At best, it’s bad (see Issue 216). Good AI detection systems — and we have to say that part — are not easy to circumvent at all. In fact, even when research teams have overtly tried to attack and trick them, it has not worked (see Issue 390). And prone to false positives? No (see Issue 433). That’s never been true. And it’s especially untrue now. Again, if we’re talking about the three to four best detection systems, which, why would you use anything else?

These are so bad, so wrong, so easy to disprove that the entire argument cannot be taken seriously. But IHE — totally cool with these falsehoods being presented as simple facts.

As for inserting her dissertation “into a popular AI detector,” name it. Ninety percent of AI detection systems are garbage, designed to give you an incorrect score. If you won’t name it, I can’t evaluate your point. It’s like saying that cars suck because you tried to move a ton of concrete in your car and it was a disaster. What car? It matters.

We also get this nugget:

The alternative is what I call a strategy of supplication, essentially trying to convince students to be responsible AI users. I see universities creating syllabus statements and online modules on AI literacy, hoping that if we explain the ethics clearly enough, students will comply. But this misses the point entirely. Students generally don’t cheat because they lack moral fiber; they cheat because they are navigating a system of incentives that prioritizes efficiency over learning.

Based on what, I have to ask, are these assertions about the reason students cheat? She does not say. IHE does not care that we don’t know or that it is, once again, opinion presented as fiat.

I mean, I agree that it’s pointless to try to convince students about the ethics of doing more work when shortcuts are so easy, so free, uninspected, and as a result, unsanctioned. But there are many reasons students cheat. Any phrase that starts that students “cheat because” is suspect. I’m not saying this is not a reason for cheating. But I am sure it is not the reason.

Moving ahead:

Last fall, I had lunch with a colleague who told me she was abandoning online teaching entirely. She’d come to love teaching online during the pandemic but felt that the pervasive use of AI had made it impossible for her to connect with students and create authentic experiences. She was especially exasperated that students were using AI to write discussion post assignments that asked for personal examples. "I ask them to share an example from their own lives, and they still give me something AI wrote,” she said, clearly frustrated.

I get it. Online teaching is the worst possible setting in which to expect integrity or genuine connection.

I’ll note as well — because I am me — that here, when we’re engaging a genuine teaching experience, we’re not getting talking points about perceived AI use. Here, it’s simply “the pervasive use of AI,” which I think most agree is far closer to accurate.

We also get:

I have concluded that the question of how to curb AI-enabled dishonesty in our classes has less to do with AI or honesty and more to do with our classes. The ease with which students can cheat using AI has exposed an uncomfortable truth: we need to do a better job teaching. We don’t need to AI-proof every single assignment or abandon teaching large online classes entirely. We do need to change the way we design and teach our classes so that the difficult work of learning, not cheating, is the more attractive option.

I loathe this argument. Always have.

I loathe that it shifts the locus of the problem from the original actors — students who absolutely know better — to teachers. Teach better. Sure. Problem solved.

And with every derisive tone I can summon — good luck with getting students to accept “that doing the difficult work of learning, not cheating, is the more attractive option.” Please let me know how that works out. The entire premise of technology is avoiding difficult work. But please, rely on telling 19 year-olds to hit the gym instead of ordering Door Dash. That will do it.

Our author shares tips that she thinks will help students do the work, instead of take shortcuts. I shan’t comment on them. But I will share this, from that section:

Human connection combats cheating, and positive social pressure is a strong motivator to do the right thing.

True. But in an online, AI-integrated, required course with 300 freshmen? Google Sisyphus.

The article concludes with this:

Many faculty see AI as a threat, both to student learning and to academic integrity. They worry that the classroom is becoming a battleground over AI ethics rather than a space for discovery. But the answer isn’t better surveillance. Instead, we need to focus on creating learning experiences that inspire students to want to do their own work. The best defense against an AI chatbot isn’t a detector or a syllabus statement; it is a class worth taking.

Don’t check. Create a learning experience that inspires. Sure. Go ahead.

Inside Higher Ed in a nutshell. Yet again.

Subscribe now

(435) How to Do Academic Integrity Wrong, Courtesy of Western University in Canada

Derek Newton — Thu, 28 May 2026 18:02:12 GMT

Subscribe below to join 5,282 (+14) other smart people who get “The Cheat Sheet.” New Issues nearly every Tuesday and Thursday.

The Cheat Sheet is verified human-written:

Western University, Professor Badly Bungle Cheating

Sent in by a reader, this story, reported by CTV, has to be the worst I’ve seen in quite some time. It’s amazing and horrifying.

It begins with cheating — at Western University (The University of Western Ontario) in Canada:

A Western University professor alleges most students in one of his classes used artificial intelligence to cheat on their final exam.

Nothing at all about that is, or ought to be, a surprise.

The professor is Jacob Shelley, who teaches as, or in, “Faculty of Law and the School of Health Studies.” So, law and health. No reason to worry about cheating there. But, according to his bio page:

Dr. Shelley is a co-director of the Health Ethics, Law & Policy (HELP) Lab at Western.

Law, health, and ethics. Because of course.

Anyway:

Shelley says he is convinced the majority of the 288 students in his health-care law course cheated on their April 24 final exam using AI.

A majority. Again, no surprise.

And when a professor says they believe — are convinced, allege — that cheating happens, I default to believing them. I trust their subject expertise and experience with students, unlike Western University. More on this in a minute.

But first:

The exam was worth 30 per cent of students’ final mark.
It was intended to be closed-book and completed independently, with no use of external sources. The exam included both written answers and multiple-choice questions.
“I had eight per cent of my class receive 100 per cent on the multiple choice. Fifty-five per cent scored over 90 per cent. I’ve never seen marks like that in 20 years of teaching,” Shelley said.

Twenty years of teaching and the professor seems pretty confident in his conclusions, or at least suspicions. CTV has video coverage of this story on the link above, which I think is worth the three minutes to hear Shelley explain some of the indications of cheating he found in this exam.

But here is where this story starts to get crazy. The professor said:

he decided not to use proctoring software because he believes it does not prevent cheating.

How’s that working?

Never mind you that this makes no sense. If watching people did not deter misconduct, no one would install security cameras. Not ever. Also, students regularly say proctoring prevents cheating (see Issue 45). But even more simply, if not using it isn’t successful — maybe it’s time for a rethink.

On point, the story quotes “tech expert and commentator Carmi Levy,” who says:

“And an unproctored exam online is almost like an invitation to use these [AI] tools,” he said.

I don’t understand how any educator does not understand this. There is far more cost to giving a remote, unsupervised assessment than there is giving no assessment at all. It’s not simply that unsupervised online assessments are useless, it’s that they’re actually dangerous and destructive because of the message they send about integrity, fairness, effort, and rigor.

To reset, we have a professor who does not secure his exams and found that, strongly suspected that, more than half his class cheated. From there, the article says that the professor decided not to count the grades from the assessment he believed as compromised:

Shelley says he decided not to use the exam results toward final grades, even though administrators instructed him otherwise — a compromise to his integrity he was not willing to make.

I’m sorry - what?

Did I read that right? That the school wanted the professor to count the grades anyway, despite that he was “convinced” there was cheating? Yes, I did:

Shelley says that in the end, and against his wishes, Western decided to include the multiple-choice portion of the exam in final grades.

I am stunned. As in, I cannot believe this.

If you set out to write a manual for how not to handle academic integrity — this is it. Don’t secure your exams, essentially inviting students to cheat. When they do, and the teacher knows it, the school ignores it and forces the grades to count anyway.

Never mind the integrity of the course, the assessment of learning, the credibility of the grade, the credit, and the degree. That’s gone. But worse, trust of honest students is gone too. Why bother doing the work? Respect from those who cheated? Gone also. They may have suspected the test was a joke, given how it was handled. But now they know for sure.

Trust and respect from your professors? You’re kidding, right? It does not get any worse than a professor who contextualizes his position in terms of his ethics and the school doesn’t care, siding instead with cheating students.

Shelly’s words:

“The message I received is that we don’t trust your judgment. We don’t care about your ethics. We care about getting marks out to students because that’s what they paid for, and we’re going to turn a blind eye to concerns about cheating,” he said.

In my view, his negligence was the first problem. But I absolutely cannot argue with this. I’m surprised he did not quit. I’d like to think I would have.

What’s true here either way is that, as he said, Western is “going to turn a blind eye to concerns about cheating.” Not going to, did.

Western is not alone. In more than a decade of covering, reading, and writing about academic integrity, my best guess is that something like 70% of schools prefer the “blind eye” approach to integrity. Which means they don’t have any.

Snapping back, let’s also consider that what the professor wanted to do here was to not count the grade — just not count it. Not assign a zero to suspected cheaters. Not initiate formal inquiries. Just not count a test he believed was compromised. If we’re to believe the reporting, the school said to count it anyway.

The school put out a statement:

Western is aware of Dr. Shelley’s claims of cheating during a final exam.
We take all measures to protect academic integrity, and we have robust policies in place to address scholastic offences.
Any student found to have committed a scholastic offence should expect appropriate consequences for their actions.
Together with instructors, Western investigates suspected instances of cheating or academic dishonesty when there is some kind of tangible evidence with which to investigate. To date, Dr. Shelley hasn’t shared any such evidence, and we continue to encourage him to do so.
Leaders at the university have been fully engaged in this matter to find an appropriate resolution for all students in this course.

I note the use of “Shelly’s claims,” which is dismissive and patronizing. I also note the school saying that they, “take all measures to protect academic integrity,” which is not true. The school allows unprotected remote assessments and overrides faculty expertise. Western, I’ll remind you here, turned off its AI detection software and in 2024, had 11 student cases of cheating with AI (see Issue 332). Eleven. The published enrollment is 42,000.

As for “robust policies,” I ask the same question to Western about that as I did to Dr. Shelley — how are those working for you? Policies without enforcement are worse than no policies at all.

I also note that the school wants “some kind of tangible evidence.” That sounds fine, but it also makes clear that the expertise of the people they hire to teach does not meet that criteria. If you watched the CTV video interview, you know how dumb that is. Putting the “prove it” on the teacher is code for “go away.”

And finally, the school is quite direct in saying they want a resolution “for all students.” Not for the professor, or for the integrity of the school, but for the students. That’s the “appropriate resolution” they want. And, it seems, the resolution they got.

And yes, most of this could have been avoided had Dr. Shelley secured his exams in the first place. Fewer students would have cheated and, for those who did, he may have been able to provide "some kind of tangible evidence,” since his word and experience were not enough.

The entire episode stinks. Of all the ways to have handled this, from start to finish, this was the worst one.

I’ll end with another quote from Levy, the technology expert:

“All schools need to figure this out and they need to do it quickly, because otherwise their very brands could be at risk of being compromised. If you cannot trust that the integrity of the degree that you grant is unimpeachable, then the value of that degree is diminished on the open market,” he said.

Not could be, are. Not if, when.

When schools, and professors, make it this clear that they simply don’t care enough to do anything about very credibly indicated cheating, when they blind eye as much as they possibly can, the party is already over.

Subscribe now

“I have an A because I use Chat”

Great reporting by Nanette Asimov from The San Francisco Chronicle (subscription required) tells us what I suspect we already know but can be reminded — students are cheating with AI.

I regret sharing so much of this outstanding reporting here. But I find it important and ask that, if you can, you visit The Chronicle and consider subscribing. Journalists and publishers need to be rewarded for work like this. It’s neither easy nor free.

The coverage starts with the groundwork — that schools prohibit cheating, that it’s up to professors to set policies on AI in their classes and remind students what is, or is not, allowed. Fair enough. And also:

reminding students they are still required to produce their own work.

True. But, sure. Good luck with that.

Numbers and Stories

Anyway, the reporting continues:

But not all students appear to be checking the rules. Conversations with a dozen of them at UC Berkeley and San Francisco State yielded story after story about how they parcel out their reading and writing to machines, leaving open the question of what they are actually learning.
“I have an A because I use Chat,” a San Francisco State psychology major said of her statistics grade. “If there’s a math problem, instead of me trying to study for a class I don’t really care about, Chat will just solve the problem for me.”

It’s not a question of checking the rules. They know the rules. They don’t care. And they don’t care because nothing ever happens when the rules are broken, ignored. The key has always been the phrase, “I don’t really care.” No reason they should.

It continues:

Two of the 12 students interviewed at UC Berkeley and San Francisco State said they abstained from AI altogether, citing the technology’s excessive use of water and energy in one case, and an unwillingness to “lose my voice” in the other. Four students said they used it only for legitimate, tutoring purposes — to explain math and science complexities or to quiz themselves — and never to do their work for them.
But six said they use AI to do some or most of their work, with some insisting that this isn’t cheating. Others said that it is. Some criticized professors for prohibiting AI, while others claimed their instructors don’t actually want them to do the assigned reading and are fine with AI-generated summaries.

Given that people don’t like to admit to disfavored behavior, especially to someone’s face, AI use for “some or most of their work” is probably more than six in 12. But even if it is not, we’re still talking about half. Just so we’re clear.

I also find it hard to believe that instructors don’t want students to read. Feels like a dose of rationalization to me. At the same time, I can see why students may think that. When you don’t read, and use AI for vapid summaries, nothing happens. Easy to see how that can be internalized as permission, or, at worst, apathy.

“I know it’s cheating, and I do feel a little oof about it,” a San Francisco State freshman said as she bit into a burrito in the student union.
She types her notes into ChatGPT then asks it to turn them into polished assignments for English and communications classes.
“It’s basically what I submit. I’ll say, this sounds a little robot-y, so I’ll tweak it a little,” she said. “But it’s getting my stuff done, so I don’t feel horrible about it.”

Still more:

At UC Berkeley, a junior majoring in media studies said she relies on AI for all assignments.
“The heavy lifting I use AI for is my reading,” she said, naming a site that turns reading material into short podcasts that she listens to before class. “Kind of like, abridged.”
What about using AI for her writing? “Yeah, 10 out of 10” assignments, she said. “Though I try not to plagiarize off ChatGPT.”
Her definition of “plagiarize,” at least, was hers alone.
She’ll have a “big sploosh of ideas,” which she types into Chat. “I’ll be like, can you condense my ideas into a few sentences? And it will. I’ll usually copy what it says. I’ll consider it mine because they’re my ideas,” she said.

Berkeley. Nice.

And yet more:

At San Francisco State, a student majoring in speech and language said she couldn’t remember the last book she had read for school, though she had recently been assigned at least three.
“We weren’t supposed to read the books,” the student claimed. “Just summarize them.”
To achieve this, she used two AI sites: one to summarize, and the second to ask whether the summaries were good enough.
The second site told her the summary of “The Glass Castle,” a memoir, needed more work.
“For that one, I did not really understand,” the student said. So she returned to the first site and told it to do a better job.
“Everybody uses” it, she said. “Since high school. It’s like a tool.”

How anyone is expected to “just summarize” while not reading the books is a mystery to me.

Bigger picture, every time I read stuff like this, I wonder whose name is going to be on the diploma and exactly what set of skills or competencies will the diploma attest that these students have? As far as I know, none of these schools offer a degree in telling the AI to do a better job because another AI said to.

Dumb Things to Say

You may think that a section titled “dumb things to say” refers to AI-using students. But, no. Not here. This section is about, and from, the folks at the schools, with a few very strong contenders for 2026 Quote of the Year.

Here’s one:

“We take the challenges seriously,” said [Ahmet] Palazoglu of the [University of California] Academic Senate. “The answer is not to pretend AI does not exist, but to be clear about what constitutes real learning. To redesign learning assessments where needed, and to have frank conversations with students about why doing the work themselves matters.”

Someone, anyone, please tell me who is suggesting that we pretend AI does not exist. And “frank conversations.” Sure. Please allow me to restate from above — the students do not care. And they won’t, unless and until you make them care.

Another:

The university must “help students see that abuse of AI’s capabilities simply undermines their ability to take full advantage of a once-in-a-lifetime opportunity,” said Oliver O’Reilly, UC Berkeley’s vice provost for undergraduate education. “When students cheat, they only cheat themselves.”

Help students see — sure. Good luck.

And if you’ve read The Cheat Sheet for any time at all, you know that “cheaters only cheat themselves” enrages me. It’s not true. Cheaters cheat their classmates, their professors, alumni, taxpayers, the school itself, and the public at large.

Yes, this is hyperbole. I cop to that. But this view has all the logical consistency of saying that people who commit crimes usually go to jail and therefore are only hurting themselves. I just can’t.

You want more? Here you are:

Pointing out that California State University has proudly gone all-in on AI, the story notes that:

CSU released “Ahead of the Curve,” a report on AI use that surveyed 80,626 students and 13,434 faculty and staff across its 22 campuses.
Nearly 80% of the students said they would not submit AI-generated work as their own. That left more than 16,000 students, about 20%, who did not check that box.
“Most students are not using AI to cheat,” said Jason Maymon, a CSU spokesperson. “The idea that AI is driving a widespread increase in academic dishonesty at the CSU or in higher education is misleading.”

A few things here that should be obvious. One, people saying they would not do something is no proxy for whether they do that thing. People say they won’t cheat on their taxes, or on their spouses, or speed in a school zone. But they do. Every single day.

And there is no way that Jason Maymon knows whether what he said is true. For the record, it’s not true. But my point is that I am sure he just made that up, that it’s based on nothing — except perhaps a desire to defend AI at CSU. The idea that AI is driving a widespread increase in cheating, at CSU and elsewhere, is beyond well established. Denying it is not a good look.

(434) Princeton to Proctor Assessments

Derek Newton — Thu, 21 May 2026 18:03:02 GMT

Subscribe below to join 5,268 other smart people who get “The Cheat Sheet.” New Issues nearly every Tuesday and Thursday.

The Cheat Sheet is verified human-written:

Princeton Changes 130-Year-Old Policy, Moves on Exam Security

There aren’t too many headline developments in the universe of academic integrity, but here is one: Princeton University will require exams to be monitored, proctored, secured.

That’s a big deal.

To dress the stage a bit, Princeton was one a few schools with what I call a ‘Capital H’ Honor Code. These are different than generic honor codes that every school has, the kind that says integrity is serious and students should not cheat. Words on paper.

The ‘Capital H’ kind says that stuff too, but it places every drop of enforcement on students. Peer reporting is the only kind. Teachers literally leave their classes during exams, actively not watching the assessment process. At Princeton this system had been in place since 1893.

But that’s over, as Princeton faculty have voted to require exam monitoring:

All in-person examinations at Princeton will be proctored starting July 1, representing the most significant change to the honor system since it was established in 1893. The faculty passed a proposal requiring instructor supervision at Monday’s faculty meeting, with one opposing vote.
The historic vote was the culmination of months of deliberation within the administration and student governing bodies about how to address increasing concerns over academic integrity violations, including the proliferation of AI usage. The proposal cleared a full faculty vote as the final of three required rounds of approval, having already been passed unanimously by the Committee on Examinations and Standing and the Faculty Advisory Committee on Policy.

This is unquestionably good. And it follows other schools with a ‘Capital H’ code moving in the same direction: Stanford University and Middlebury College (see Issue 209 and Issue 312).

It is odd, though, that according to the new policy, teachers at Princeton won’t be able to stop cheating should they see it. They can only witness it and report it to the student integrity committee, as a witness. Still, added eyes will be a deterrent:

“Undergraduates and faculty are realistic in understanding that having an instructor supervising examinations will not eradicate cheating,” the proposal notes. “However, they believe that there will be a significant deterrent effect, and that having an additional witness in the room will reduce pressure on students to notice and report concerns while they are themselves completing exams.”

What’s most interesting about the moves at Stanford, Princeton, and Middlebury is that, to a large extent, the shift has been driven by students.

Turns out that two things are true. One, honest students really hate when their peers get away with taking shortcuts, getting better grades without doing the work. And two, as cheating increases, students also really hate being the only ones responsible for dealing with it, being expected to turn in their classmates as the only remedy. So much so, that they’re not doing it. At all.

‘Capital H’ approaches were not working:

In The Daily Princetonian’s 2025 Senior Survey of over 500 seniors, 29.9 percent of respondents reported that they had cheated on an assignment or exam during their time at Princeton. 44.6 percent of senior respondents reported knowledge of Honor Code violations that they chose not to report. Only 0.4 percent of seniors responded saying that they had reported a peer for an Honor Code violation.

From the new policy:

If students alone are present in the examination room and students are unwilling to report, then there is no check against misconduct during assessments.

There is the issue — no check against misconduct.

Here’s more, from the approved change itself:

Over the past few years, and with increasing frequency over the last six months, significant numbers of undergraduate students and faculty members have requested that the Office of the Dean of the College consider a policy change toward proctored examinations, given their perception that cheating on in-class exams has become widespread.

Significant numbers of students asked for the change because they believe cheating has become widespread, to paraphrase. They’d know.

We should also put on the record that AI was a driver in this change. From the approved policy:

Commonly cited are the advent of generative artificial intelligence products which significantly lower the barrier to gaining unfair advantage in the context of an in-class examination.

AI lowers the bar to unfair advantage. To paraphrase again.

Of note, the new policy does not allow proctoring of in-class exams, it requires it. Also worth noting that it applies only to in-class assessments. No word on how — or even if — Princeton is monitoring remote assessments, which are far more vulnerable to misconduct. Whether the policy opens the door to online exam proctoring, for example, is unclear. It may not.

But the biggest news is that over the past few years, a handful of highly visible ‘Capital H’ Honor Code schools have chucked that policy, choosing to make their assessments more secure, improving the rigor of their instruction, and investing in the understood quality of the grades, credits, and credentials they provide.

As much as I lament and complain about schools that continue to force their eyes closed on issues of fairness and integrity, there are schools that have seen the problem and actually moved to fix it. They deserve a ton of credit. As do the students who have pushed for more help, more accountability, more equity.

Subscribe now

Guest Essay in NYT

Recently, the New York Times ran a Guest Essay from a senior and book author at Stanford University. The headline, “What A.I. Did to My College Class.”

The piece is a mish-mosh, in my opinion. Kind of all over the place. But whatever. It has a few tidbits about cheating worth sharing, especially considering they were in the Paper of Record. Here is the setup:

We are the first college class of the A.I. era — ChatGPT arrived on campus about two months after we did.

The essay also says, pointing to several high-profile business fraudsters:

Stanford already had a shaky reputation for integrity when I arrived in 2022.

And:

Now A.I. has made deception easier and more remunerative than ever before.

The context makes it clear the author is talking about business, not academics. But this is on point:

Cheating has become omnipresent. I don’t know a single person who hasn’t used A.I. to get through some assignment in college, yet the school was at first slow to realize how widespread this would become. As freshman year went on, some professors suggested that the “nuclear option” might be called for: allowing faculty to proctor in-person exams, a practice banned at the university for over a century to demonstrate “confidence in the honor” of students.

This ‘nuclear option’ happened.

There’s also this:

In junior year, 49 percent of the 849 computer science majors who responded to an annual campus survey said they would rather cheat on an exam than fail.

About halfway through freshman year, some coding classes started requiring students to sign a declaration — “I did not utilize ChatGPT” — to submit each assignment. During the first term these attestations began to appear, I watched a freshman I knew sign the declaration that he’d done his homework without A.I. as ChatGPT was still open in the next window — while on the deck of a yacht party financed by venture capitalists. The incentive structures were not aligned toward honesty. One could get ahead, quickly, by cutting corners, by focusing on self-presentation.

As I’ve said before, these self-declarations are useless. People who cheat lie about cheating. I guess these statements do no harm. But they do little good, if any.

Want more? Here you are:

Students were probably the earliest wide-scale adopters [of ChatGPT]. After all, it was far and away the quickest route to an A. When I took CS107, the only viable way for people to cheat was to seek out a student who’d gone through the class before and beg for solutions to the notoriously difficult problem sets. There was no alternative to putting in a large amount of work. Even if one did obtain the answers from another student (engaging, by the way, in a social act, if nothing else), the students I knew who did this still spent hours sculpting their stolen code so as not to be caught.
Few cheated in this most overt fashion back then. But a month later, any student could instead turn to a chatbot, plugging in a prompt alone in a dorm room and mindlessly regurgitating the result. “I remember the first time I used it feeling an immediate sense of guilt,” a friend recently told me. “Now it’s just normal.”

Just normal.

It continues:

Half of the laptops in any lecture seem to be open to ChatGPT or Claude.

This is good too, I think:

Emerging research has begun to show what most people feel is obvious: Relying on A.I. for cognitive tasks can reduce one’s own intellectual capacity and resilience. It’s one thing to use it in the workplace, but in the classroom, difficulty is often precisely the point. Sure, a robot can lift 600 pounds much more easily than I can — but that doesn’t much help me if I’m trying to work out. The same goes for the thinking exercise of education. However, telling that to students is about as attractive a message as “eat your veggies” or “sleep eight hours.” It feels like scolding.

I don’t have much to add. So, I won’t.

Subscribe now

From Studiosity: To get back to learning, move past surveillance

A week or so ago, I had the chance to see a demo of a new learning management and assessment system from Studiosity.

After the demo, I invited Studiosity to submit something for The Cheat Sheet. Space here is always open, for free, to anyone or any company with anything to share about academic integrity.

Here is what Studiosity sent, unedited:

Helping universities move beyond ‘detection’ as a flawed strategy, Studiosity is leading a different approach to assessment security; where students’ critical thinking skills and authorship verify graduates.

In practice, the Studiosity platform integrates authorship validation directly into standard LMS assignment setup, without the need for post-assessment ‘policing.’ Educators keep their current assignment workflow. They select validation thresholds upfront (the platform offers frameworks like the AI for Assessment Scale, Perkins et al.). Students submit as usual, and self-validate through adaptive viva-style questions and cloze testing. At the same time, the platform considers authorship continuity and develops students’ critical thinking in their writing, over time.

Because choices are made during assignment setup, educators receive clear “valid” indicators after cohorts submit, without needing to brace for ongoing detective work. For college leadership, it’s scalable, pedagogy-backed infrastructure to make those new AI policies actionable, and to finally defend graduate skills - for any level of AI use by students.

Colleges in North America, UK, and Australia will use Studiosity’s assessment security platform this July.
College leadership should inquire: studiosity.com/validate

Subscribe now

(433) Law Stuff and Cheating with AI

Derek Newton — Tue, 19 May 2026 16:36:23 GMT

Subscribe below to join 5,268 (+15) other smart people who get “The Cheat Sheet.” New Issues nearly every Tuesday and Thursday.

The Cheat Sheet is verified human-written:

San Francisco High School Student Sues over AI Misconduct Accusation

The San Francisco Standard has the story of a Palo Alto high school student who was accused of using AI to cheat and has sued his school.

From the article:

In October, a sophomore submitted an essay on the text to his English teacher at Palo Alto High School, Sarah Bartlett. By December, he had been accused of using mostly AI to compose it and was required to do an in-class rewrite — which got him a D.

Later in the coverage we learn that the original essay, the one the student was accused of using AI to complete, received a grade of B. To me, it says something that, when asked to do the assignment again, by hand, the student earned a D.

Here, it seems that at least part of the consequence of being suspected of using AI on an assignment was to do it again. That does not feel like a significant penalty, to be honest. So not serious, in fact, that if the punishment is a do-over, what’s the risk in trying to shortcut the first attempt?

But most importantly, I infer that the school or the teacher did not escalate a suspicion of misconduct to actual sanctions. At least not right away. That feels important.

Responding, and Turnitin

Continuing:

[The student’s] family submitted drafts, timestamps, and direct access to his Google Doc revision history — culminating in a 1,162-page evidentiary packet sent to the school in January. By March, the family made an offer to the school: Give the kid a B, and everyone goes home.

1,162 pages. So, here we go.

Also, a Google Doc revision history is not necessarily evidence of not using AI. It can be faked (see Issue 427). But seriously, 1,162 pages. To rip The Bard, they doth protest too much, methinks.

By the way, Turnitin’s AI detection tools, used by the school in this case, flagged his essay as 76% AI, according to the article. That’s not some insignificant rounding error.

Anyway, the school declined the “let’s pretend the AI thing never happened and give him the original grade” offer. So, naturally, the family is suing.

Bigger Picture

The article continues:

It is a stress test of a system that schools across the U.S. have quietly adopted with almost no guardrails: AI-detection software that its own maker acknowledges is fallible, deployed by individual teachers with no standardized threshold for action, no required review process, and no reliable way for wrongly accused students to clear their names.

I’ll grant that AI detection software is deployed by individual teachers with no standards. And that’s a real problem. Schools should, at a minimum, be sure that teachers are using the same system, a good system, across all students and assignments. And, I also grant that it is hard for accused students to clear an accusation. Fair. Though, a re-test without access to AI seems like a decent way to do it.

On the other hand, a standardized threshold for action is a bad idea (see Issue 396). And, the fallibility of AI detection systems is, as we have pointed out repeatedly, wildly overstated. It is, as I have also said before, silly.

False Positives are Silly

They are silly because we seem to want AI detection in schools to be 100% accurate, all the time. That’s not a standard we apply anywhere else. Here are some examples from Google’s AI search on reliability and false positive rates from some other, common, everyday detection and alert systems:

Airport X-ray Baggage Scanning (EDSCB): Modern Automated Explosives Detection System for Cabin Baggage (EDSCB) of Standard C2 produces automation false alarm rates of 6%–20%.
CT-Based Scanners: Advanced CT-based security systems can deliver a false alarm rate as low as 5%.
Perimeter Detection Sensors: In industrial security, outdoor perimeter sensors are prone to nuisance alarms, with some studies showing high false positive rates of over 20% depending on environmental factors.
Automatic Fire Alarms: Reports show that up to 98% of automatic fire alarms are false alarms.
ICU Monitors: Studies in hospital intensive care units indicate that 80%–99% of alarms are false or clinically insignificant.
Radar/Sonar: Without optimization, false alarms can be high, but optimized, specific methods can bring false detection rates down to 4%–18% in sonar imagery.

Research has shown that 77% of all mammogram alerts are false positives — more than 3% of everyone who gets a mammogram gets a false positive.

And yet, I’ve never heard anyone argue that, due to their high rates of false positives, we should discontinue using airport baggage screeners, or ICU alarms, or fire alarms, or mammograms. We continue using all those, even highly encouraging their use, because it’s really important to catch the serious, dangerous stuff. In these examples and hundreds more, we’d rather have 50 false alarms than miss one serious threat.

We simply don’t care enough about academic integrity and fairness to apply standards that are anywhere near similar.

When it comes to teaching and learning, we’d rather let a generation of students cheat to their pleasure than tolerate an even 1% false alarm rate. We simply don’t deem cheating to be a serious enough threat to do the work of investigating an alarm, potentially dismissing one in a hundred as inaccurate. We prefer not to know, to let it burn.

Not all of us, obviously. But plenty, nonetheless.

We could, of course, decide that academic rigor, the value of academic credentials, and basic fairness dictate that we adopt a similar mindset to the one we use for nearly every other detection system on Earth — doing the work to investigate and perhaps dismiss a few false alarms because what we’re trying to find is actually important. But too many education leaders have decided not to. They have decided that a 1% false positive rate at the review phase — not even at the investigation or accusation phase — is disqualifying.

That’s a choice.

Rant Over, Back to Palo Alto

Anyway, on the topic of triggering further inquiry, and back to the Palo Alto case, the school may run into trouble here:

The suit alleges that the incident reduced Kato’s semester grade from a low A or high B to a C. His parents were not notified, as school policy outlines, according to the suit.

Schools must — absolutely must — follow their policies.

Still, if they don’t, when they don’t, it in no way means there was no misconduct. It also in no way indicates that the AI detection system failed.

Discrimination and Bad Reporting

The suit alleges, according to the coverage, discrimination based on race and gender:

The complaint alleges that at least one other Asian student in Bartlett’s class was subjected to the same treatment — flagged by Turnitin, forced to rewrite in class, and handed a D that gutted a previously strong grade. The complaint also alleges a gender disparity: that male students in Bartlett’s class were four to five times more likely than female students to be subjected to the Turnitin assessment and punitive grade replacement.

On the allegations of discrimination, I cannot speak to the issue of Asian students being targeted. But I can say that, in general, males tend to cheat more, to try to cheat more often. So, any fair application of inquiry and consequence would tend to include more male students than female students.

At the same time, this is bad reporting, a trend that gets worse as the story continues.

Note please the use of “subjected to,” and “forced to,” and “handed a grade that gutted.” And also, “punitive grade replacement.” About which, being asked to replicate an assignment in an alternative format is not necessarily punitive. It’s just as much an opportunity to demonstrate mastery, to show, in other words, that you didn’t use outside resources inappropriately or unfairly.

After that, the article drags us through other supposed cases of students being wrongly accused of AI use:

The Palo Alto case is not an isolated one. Leigh Burrell, a sophomore at the University of Houston, received a zero on an assignment worth 15% of her final grade after her professor suspected she had used AI — despite Google Docs history showing she had drafted and revised the work over two days. Students at the University of Buffalo started a petition after Kelsey Auman, a master’s student, had her work flagged by Turnitin at the end of the semester and was threatened with not being able to graduate.

I’m not saying Leigh Burrell cheated. I am saying Google Doc history, as previously mentioned, can be faked. And students at University at Buffalo started a petition. Stop the presses. The petition, by the way, is a joke (see Issue 365).

The article also slogs through this again:

At the University of Minnesota, a third-year Ph.D. student named Haishan Yang claims he was expelled after faculty compared his exam answers to ChatGPT output, costing him his legal right to remain in the United States.

If you care, please see Issue 336. Forget the AI thing. In a human review, a panel of professors decided the submitted work was not his own.

Not to stop, the piece also incorrectly rolls through a few studies allegedly showing that AI detection in general, and Turnitin specifically, are flawed:

A 2023 study published in the International Journal for Educational Integrity tested 14 popular AI detectors and found that even the best-performing tool — Turnitin — struggled significantly with paraphrased AI content, with more than 50% of manually edited AI text going undetected across the tools tested.

This is the Weber-Wulff study we covered in Issue 250, which is now three years old. But in citing it, the reporter singles out “paraphrased AI content,” and “manually edited” text — which we’re not even talking about in this Palo Alto case. And about AI text “going undetected” by Turnitin, which, also is not the issue here. This Palo Alto deal is supposed to be about a false positive — incorrectly flagging human writing as AI-generated. Different.

Had the writer of this news article bothered to read that study, or our coverage of it, he would have seen that the 14 tested AI detectors were 96% accurate at detecting human-written text — not flagging human writing as AI, in other words. Turnitin was perfect. The story cited a study to discredit Turnitin that actually showed it performing perfectly.

That’s bad.

The article cites this as well:

Stanford University research indicates that neurodivergent students and students for whom English is a second language are flagged at significantly higher rates than others, because their reliance on repeated phrases and consistent terminology mimics patterns the algorithm associates with AI output.

Again, not at all related to the case at hand. But also, this study has some serious errors and issues, which we covered in Issue 216. One, which I will just keep screaming about, is that the paper has completely made-up citations. No one cares. We just keep repeating it, even where it is not related to the topic — like here. Because, you know, AI detection is bad.

And Cash

Finally, the reporting says:

The family is asking the court to vacate the rewrite grade, restore Kato’s original semester result, and order the district to expunge any reference to academic dishonesty from every category of record. They are also seeking compensatory damages and attorneys’ fees, and asking the court to bar PAUSD from treating AI detection scores as dispositive evidence without educator review and a meaningful opportunity for the student to respond.

They want the grade the student received with AI access, the one for which he may have — likely did — use AI. Plus, they want money. And to stop the school from using AI detection without educator review, which is fine, although it’s very unlikely that this case resulted from an AI detection score alone.

And, Grammarly

Finally, there’s also this:

The teacher’s justification for the punitive rewrite rested on the Turnitin score and what she claimed was Kato’s admission to using Grammarly for “synonym assistance.” The Kato family denies the admission ever happened. But even if it had, Grammarly is a grammar and editing assistant that millions of students use routinely, without thinking of it as “AI” at all.
There is another, largely unreported, problem: According to Turnitin, the use of Grammarly can trigger higher AI detection scores. A 2025 study in the International Journal of Educational Technology in Higher Education found that even a known human-written abstract that was edited with Grammarly returned a 16% AI score on Turnitin.
Grammarly acknowledges that its more advanced rewriting features — including paraphrasing and proofreading agents — are powered by large language models and “should still lead to some percentage of AI-generated text being triggered in AI detection.” Even Turnitin’s own educator forum community has flagged the problem, with teachers noting that student work composed using Grammarly can return AI scores of 100%.

First, again, a rewrite is not punitive. But, Grammarly. Because, of course.

I beg you to read this sentence again, from above:

But even if it had, Grammarly is a grammar and editing assistant that millions of students use routinely, without thinking of it as “AI” at all.

I mean, sure. Grammarly is used by millions of students who do not think of it as AI. But it is. And that someone uses it “without thinking” does not justify the “But,” which starts that sentence. I may have been speeding, your Honor — but millions of drivers do this routinely without thinking of it at all.

Not a compelling argument.

And yes — using Grammarly can return AI detection scores because (wait for it) Grammarly is AI. This is not, as the reporter says, “a problem.” This is AI detection. Grammarly is AI. It is being detected. How is this complicated? Did the reporter not read his own writing:

Grammarly acknowledges that its more advanced rewriting features — including paraphrasing and proofreading agents — are powered by large language models and “should still lead to some percentage of AI-generated text being triggered in AI detection.”

Grammarly is AI. It is. And it will get flagged. Literally everyone agrees on this. Even Grammarly.

And, back to our Palo Alto issue, our student may have used it. Raise your hand if you’re surprised. My hand is not up.

We’re going to keep seeing this — students accused of cheating with AI, suing the school to get the results thrown out. A few of those, as we’ve already seen, will “succeed” because the schools did not follow their own policies (see Issue 428). To me, these cases are worth tracking very closely, if only because those who don’t want to bother with cheating, or those invested in convincing schools not to “police” AI use, will cite them as support for their position.

Subscribe now

When AI Cheating Becomes a Legal Risk

Continuing a theme, our friend, and legit expert in academic integrity issues, Christian Moriarty, has penned an article at Chronicle of Higher Ed about the legal risks, from a school’s perspective, associated with AI cheating.

The entire piece is good.

But, for ego issues, I’ll start by sharing this, related to the case cited in the earlier story — the one above, about the PhD student in Minnesota who was expelled. There, the case is cited as some kind of warning about AI detection and consequences. Here’s what a real expert wrote:

Consider what happened at the University of Minnesota in 2024. Haishan Yang, a third-year Ph.D. student, took a remote qualifying exam. Faculty compared his answers to ChatGPT output and found similarities. The university expelled him, costing him his doctoral program, his student visa, and his legal right to remain in the United States. Yang sued, alleging violations of due process. Both a federal district court and the Minnesota Court of Appeals upheld the expulsion, finding the university’s decision reasoned and evidence-based. That record, built through notice, a formal hearing, advocacy, and appellate review, gave the institution a basis neither court could characterize as irrational. That is the floor: The severity of the consequence demands the construction of that record, and without it, the institution is exposed.

The expulsion was upheld. Twice.

Moriarty compares that Minnesota case with the examples from Adelphi University, in which a student sued and won, due largely to the school’s failure to consider evidence and a meaningless appeals process (see Issue 428). He writes:

institutions watching have no way to predict which quiet zero will become the next lawsuit. The only question is whether institutions have any framework governing decisions about how to handle misconduct. Those that do will keep their students answerable to them; those that do not may find themselves answering to judges instead.

Framework. Having it — and following it — is the difference.

Importantly, Moriarty also distinguishes between grade outcomes and disciplinary ones. Grades, he writes, are usually left to teachers. On matters of discipline, on the other hand:

Due-process considerations apply: notice of the allegation, an explanation of the evidence, and a meaningful opportunity to respond. At private institutions, contract law typically imposes a parallel obligation to follow published procedures.

He continues:

The difficulty is not only that institutional norms haven’t caught up, but also that AI detection produces predictive evidence that resists the binary findings misconduct proceedings require, and the algorithms that generate those predictions aren’t available for the kind of cross-examination a fair process assumes.
Do AI-detection tools produce false positives? How many? How do you know this isn’t one of them? Can you explain how the detector arrived at this result? These are elementary cross-examination questions, and you cannot answer them because the detector is largely a black box. That uncertainty alone creates real doubt under the preponderance standard.

These are good notes, although I am not sure they are accurate in a factual sense.

Of the handful of good detection systems, the companies that built and run them can answer these kinds of questions. They can tell you why a particular piece of writing was flagged. They can tell you how a result was calculated. They can tell you how they know, to what degree of certainty, that a particular document is not a false positive. In the context of a disciplinary inquiry, I just don’t think schools ask.

In fact, an aside, I’ve suggested to a few AI detection companies that they offer a forensic-level report for formal inquiries, as a service — even offering to explain or answer questions about it in a hearing. For a fee, of course.

But I continue to suspect that, even if such a thing was available, schools would not use it because they don’t want to pay, and because they don’t want to be seen as escalating the hearing process. Although — news update — the process is already escalated. Entire law firms exist to represent students accused of cheating. And the legal challenges do come.

My point is that AI detection probably is a black box to teachers, likely doubly so for administrators. But AI detection is not a mystery. It’s math. The people who build AI detection systems, update them, hone them with intention, know them very well. Usually, I’ve found those folks are downright eager to open these boxes, if anyone asks.

But also, none of this matters because there is wide agreement that scores from AI detection software should be supplementary evidence, not sufficient evidence. No one should be bringing a formal hearing of misconduct based on an AI detection score alone, let alone substantiating one. Personally, give me an AI score from a good system and a teacher’s assessment, that’s good enough. Just one of those, maybe not.

Another strong insight from the piece:

If the consequence remains within a single assignment, a professor applies reasonable standards for originality, citations, and compliance with the course policies and assigns a grade. That is a faculty decision reviewable only through standard grade appeals. Courts will likely defer to it.
If the consequence extends beyond that assignment — to course failure, a transcript notation, suspension, or expulsion — the institution has entered disciplinary territory. The student is owed due process, regardless of what the professor calls it or how the institution categorizes the proceeding.

Frankly — my opinion — due process should not be that hard to deliver.

Finally, Moriarty writes:

Institutions are not helpless here, as legal and moral frameworks exist. What is missing is the will to apply them consistently, and the administrative clarity to distinguish between a faculty member exercising professional judgment about a piece of work and an institution making a formal finding of misconduct

He’s right. More even than he meant, perhaps.

On academic integrity and fairness, we’re missing the will to do just about anything.

Subscribe now

(432) Cheating at Purdue University and the Jaw-Dropping Response

Derek Newton — Thu, 30 Apr 2026 17:51:51 GMT

Subscribe below to join 5,253 (+0) other smart people who get “The Cheat Sheet.” New Issues nearly every Tuesday and Thursday.

The Cheat Sheet is verified human-written:

An Unbelievable “Cheating Scandal” at Purdue

Local news has the story of a “cheating scandal” in the computer science department at Purdue University.

Despite the nice narrative writing, here is where the story starts:

Computer science professor Jeffrey Turkstra sent an email to more than 200 students in his CS 240 class last week, alleging he had caught them cheating on assignments with AI tools. He included a form questioning AI use on each assignment, giving accused students a chance to admit or deny the allegations.
“Failure to respond will result in a grade of ‘F’ for the course,” the email reads. “Failure to respond will also result in an unfavorable letter being sent to the dean of students along with further disciplinary action.”
Sent on the deadline for dropping a class, students went into a frenzy, and more than half the class unenrolled.

The news reporting also described the fallout as “chaos.”

I’m not sure whether sending this e-mail on the drop deadline was by design, but it says a ton to me that more than half the class quit — not just that the professor may have been right, but that a culture exists at Purdue in which more than half the students may have felt empowered to cheat in the first place.

Continuing, from the coverage:

The majority of students in the class, including many who admitted online to using AI, are not upset that potential cheaters were caught, they said. They’re upset about how it was handled.
“The main issue that almost everybody has with this is the way that Turkstra went about it,” one student, who asked to remain unidentified to protect his grade, told the Journal & Courier. “The way he tries to fear-monger us without telling us which ones we used AI on, his detection methods, his timing.”

Many admitted to using AI. I want to be sure we don’t lose sight of that. I also don’t think the professor owes it anyone to explain his methods, timing, or individualized suspicions, especially in a group setting. One-on-one, or in an integrity hearing, sure.

Either way, a main complaint of the professor’s process was that Turkstra was required, or maybe just expected, to meet with each student individually to discuss his concerns, which did not happen — at least not before his mass e-mail.

Plus, if my math is right, we’d be talking about more than a hundred student meetings. How’s that supposed to work?

Even so, this probably was not required anyway. From the article:

But it’s unclear whether the informal meeting — which Turkstra seems to have bypassed — is required or merely suggested.

And:

The Office of Student Rights and Responsibility’s academic integrity page says the meeting is “encouraged,” leaving possible leeway for professors.

So, leeway.

But then there’s the other shoe:

The CS department head sent an email over the weekend announcing that Monday’s lecture would be a “class discussion,” to hear concerns and create a plan moving forward.
It seems they unofficially ruled in students’ favor, with Monday’s “discussion” amounting to a big “never mind.”
Turkstra told students that after talking with department heads, he understood the timing could be seen as “coercive,” though that was not his intention, the unidentified student said.
Monday morning, it was announced that allegations of cheating and AI use thus far in the semester are dropped, and anyone can re-enroll in the course.

The department intervened and all cases were dropped. Cheating? Don’t worry about it — we got you. A “big never mind.”

Not to put too fine a point on this, but we have every reason to believe that a majority of students in this class were shortcutting their learning, violating integrity policies, and, more likely than not, in position to receive grades, credits, and degrees they did not earn.

Purdue’s response: all good.

I wish I could say this was an outlier. But Purdue’s history with integrity issues has been well documented, including, unbelievably, partnering with cheating provider Chegg to run the school’s writing center. No, really (see Issue 324). And Purdue has known about cheating in its CS and engineering programs for a long time, for at least as long as I’ve been writing The Cheat Sheet (see Issue 8). That was 2021.

Nonetheless, this one is hard to top.

If you’re actively looking to feed a culture of misconduct at your school, this seems like a solid way to do it. If you want to destroy trust and the dynamic of support with your teachers, this checks that box too.

I mean, in my most incredulous tone, this is unbelievable. If the reporting is right, a professor reasonably suspected a big chunk of his class was bypassing learning with disallowed shortcuts — cheating, to be clear. More than half the class quit instead of answering for their actions. The professor’s boss called timeout, stepped in, and made it all go away.

Skeptical me says that having more than half the students drop the class was what motivated the administrators. Purdue, like many schools I am sure, would prefer cheaters pay their tuition. After that — whatever.

It’s probably also not a coincidence that, as the paper also reports:

Purdue has been on the front lines of embracing AI, going so far as to add an AI literacy requirement for graduation.

Pretty awkward to hold students accountable for using AI when the school is “on the front lines of embracing AI.”

But at least there’s one sane person at Purdue:

Turkstra said he’ll still use the AI detection tool going forward.

That’s good, assuming the school will let him do anything with what he finds.

I’ll end with this:

Most people invested in the situation seem to agree that using AI to write code is cheating and fundamentally diminishes the learning experience.
But they also agree that AI does have a place in the classroom, even if it’s not clear exactly how yet.
“What the university should be doing is teaching us how to use the tool,” one student said, “instead of slapping us on the wrists every time we try.”

Name one wrist that was slapped here — other than the professor’s. Go ahead. I’ll wait.

Instead, even when the professor has a no-AI policy, even when the professor looks, and acts — nothing. I don’t have a better word than unbelievable.

Subscribe now

Webinar: AI Agents and Learning

From our friends at This Isn’t Fine, the outstanding action-focused newsletter on academic misconduct, is this Moodle webinar (below). Seriously, all we do over here is complain about stuff.

This isn't fine

Moodle.com Webinar: AI Agents and Learning

Moodle, the largest Learning Management System by users and sites, is hosting a webinar on Tuesday, May 5 at 12:00 PM EST focused on AI Agents…

2 months ago · Joseph Thibault

The event is this next Tuesday, May 5th at noon ET. If you’re a professor or administrator, sounds as though it may be worth checking out.

Subscribe now

Department of Corrections Department

In the last Issue, I listed the authors of a new paper on AI use in academic settings:

Sina Rismanchian, Peter Liu, Gabe Avakian Orona, Duncan Pritchard, and Shayan Doroudi

Then, in noting that Orona was affiliated with Boston College, I misspelled Orona. It was right there. I just goofed. I apologize.

Subscribe now

(431) Research: More than Half of Students Who Use AI on Assignments Do Not Admit It

Derek Newton — Tue, 28 Apr 2026 19:01:34 GMT

Subscribe below to join 5,253 (+26) other smart people who get “The Cheat Sheet.” New Issues nearly every Tuesday and Thursday.

The Cheat Sheet is verified human-written:

Research: Self-Reported Data on Misconduct Likely Inaccurate

New research reinforces the accepted framework that surveys about academic integrity, especially those in which students report on their own conduct, including AI use, likely under-report actual behaviors that may be disallowed or disfavored.

As examples, Issue 52 and Issue 276 include research showing that actual illicit or disfavored conduct rates may be multiples higher than self-reporting indicates — something to consider when we see self-report survey data showing that 75% of college students admit to cheating (see Issue 138).

Set-Up and Limitations

The paper is by:

Sina Rismanchian, Peter Liu, Gabe Avakian Orona, Duncan Pritchard, and Shayan Doroudi

Orana is affiliated with Boston College. The others affiliate with the University of California, Irvine.

The study examined the online work of 81 undergraduate students in “three large general education philosophy courses” in the United States, looking for:

Concerning AI Usage (CAI) (i.e., using AI to complete all or much of an assignment with minimal input from the student)

CAI is, according to the paper:

empirical indicators of AI involvement, including embedded content markers and formatting artifacts linked to AI tools, coder-judged signs of likely AI-generated text, unusually rapid completion times, and evidence that responses were pasted rather than typed.

And:

detailed Canvas action logs, including timestamps, typing duration, paste events, and autosave activity

The team used “Trojan Horse” detection and/or data related to time-on-task — both of which are quite solid evidence of AI use, or other non-scholastic effort.

There are some caveats to this paper.

As the authors share, the sample size is on the small side. Additionally, the team’s approach to uncovering AI is robust, but not complete — some AI use may have escaped detection in this study. Further, the assignments under analysis in this paper were for extra credit in these courses, not grades. And no specific rules or reminders regarding AI use were expressly given about these assignments. Finally, this paper examined AI use and misuse only, not other kinds of misconduct.

The caveat about AI rules and guidelines, fine. Part of the paper’s experiment was to see to what extent self-regulation impacted incidents of what should be obvious shortcuts. The extra credit assignments were about “intellectual virtues” and included self-reporting on aspects of the topic. And as the authors say throughout, citing corroborating research, students know when AI use and other shortcuts are unethical.

That the work was extra credit and not part of the class proper is interesting. If you believe, as I tend to, that low value work is more likely to be compromised by shortcuts and fraud, the nature of this study may overstate misconduct rates we may find in more normal, more consequential course settings.

Conversely, if you lean to the idea that stress and pressure to compete drive a significant portion of cheating behavior, as many do, then you might take these results as actually understating the cheating.

Findings

Either way, the paper’s core finding is that:

depending on which signals we rely upon to detect AI usage—41% to 70% of students engaged in CAI.

The 41% were students whose assignments showed multiple markers of AI-enabled shortcuts. The 70% was students with at least one such marker. But also:

Alarmingly, we found that 19% of our students were flagged for Possibly or Very Likely CAI on all seven assignments that they were supposed to complete.

If you believe, as most do, that offloading academic work to AI does not produce learning — 41% of students were not learning on at least some of this coursework. And that rate may be as high as seven in ten, with nearly one in five turning to AI for every assignment. And, as mentioned, this could actually under-report the rates in actual, grade-granting, credit-granting, degree-granting settings.

But the paper’s more important finding, in my view, is that students who used AI to take academic shortcuts and avoid learning, lied about it. Shocker, I know.

Consider, from the paper:

our findings indicate that students who engage in Concerning AI Usage often conceal these behaviors in self-reports. The discrepancy between observed behavior and self-reported accounts highlights the limitations of survey-based measures for assessing AI-related misconduct.

And:

In fact, more than half (52%) of students who showed behavioral signs of concerning AI use reported that they had not used AI at all, underscoring the limits of self-reported data.

More. Than. Half.

I don’t know what to say that the majority of students who used AI to offload work on an extra credit assignment were dishonest about their dishonesty. Seems obvious, although the number should give most people pause. I hope. And, I also hope, something to consider the next time we see reports of suspected AI use where the student denies it — which we will no doubt see again before I finish this sentence.

The authors say:

Headlines and institutional narratives are often grounded in anecdote or isolated cases, while most empirical studies rely on self-report surveys, interviews, or perception-based measures. Such approaches are vulnerable to social desirability bias and provide limited insight into students’ actual behavior.

And, citing similar research:

These findings suggest that students may systematically underreport CAI when asked directly.

To summarize, I’ve got three key takeaways from this paper:

Use of AI for academic shortcuts is common
Self-reporting of misconduct is bogus
Students use AI and lie about it

Also Noteworthy

There are two other findings worth passing along.

The team found that the link between high “self-reported personal integrity and curiosity” and CAI (Concerning AI Usage) was backwards. From the paper:

for curiosity and personal integrity, we found that there was a significant relationship with CAI, in the opposite direction of what we hypothesized, meaning that students with higher self-reported personal integrity and curiosity had a higher probability of engaging with Concerning AI Usage behaviors

Interesting. Pure speculation — this may be because this subset of students knows what the “right” answer is but really does not care. I don’t think the paper mentioned, or whether the sample size is robust enough to illuminate this, but I’d bet that the overlap between the set of high self-reported personal integrity students and the set of students who used AI but did not admit it, is substantial.

Finally, it’s worth noting that the team asked students to write answers that were both “knowledge based” and “personal-type,” based on their personal experiences and views. The team found no difference in AI use between these two approaches:

Personal-type questions did not differ significantly from knowledge questions

And:

20% of students used AI at least once to generate personal anecdotes, a practice that directly conflicts with integrity because the response is likely not genuinely about the student.

I’ve heard many, many educators and pundits share that switching assessments to include more personal, reflective, and experience-based questions is somehow a deterrent to AI use. It’s not. The AI can make up personal experiences very easily. More importantly, the only way a teacher or professor would know if a “personal” experience is invented is by knowing their students, which, we can all admit, is quite rare.

Anyway, students are using AI to shortcut the work of learning — probably more frequently than our baseline, survey-based data shows. And most are not being honest about AI use.

For the record.

Subscribe now

The Podcast of Record

If you’re not familiar with it, or not already following along, The Opposite of Cheating podcast is becoming, may already be, the podcast of record related to academic integrity and cheating.

Hosted by Dr. Tricia Bertram Gallant and Dr. David Rettinger, two highly credible voices on the topic, the show is somehow up to 58 episodes already. And though I will never catch up, I know enough to recommend it. If podcasts are one of your things, check it out.

Subscribe now

(430) AI Misconduct Cases Surged at University of Florida

Derek Newton — Thu, 16 Apr 2026 15:52:01 GMT

Subscribe below to join 5,227 (-3) other smart people who get “The Cheat Sheet.” New Issues nearly every Tuesday and Thursday.

As always, The Cheat Sheet is verified human-written:

AI Misconduct Cases Continue to Climb at University of Florida

According to coverage in the student paper at the University of Florida, AI-related academic misconduct cases at the school “have surged in recent semesters.”

Before we get into it, a few points of privilege. We continue to owe a significant debt to student reporters and student papers, which are often our single source of information on academic integrity. Also, this story on cases of misconduct has this at the top:

Read other stories from the "These stories were not AI-generated" special edition here.

I think that’s pretty cool.

Here are the first few paragraphs:

Reports of academic misconduct at UF that reference artificial intelligence violations have surged in recent semesters, according to university records obtained through a public records request.
From Fall 2021 to Fall 2023, UF reported no Honor Code violations that included terms like “artificial intelligence,” “AI” or “ChatGPT.”
In Spring 2024, 20 cases were identified; that figure rose to 42 in Fall 2024 and 66 in Spring 2025. The records only count misconduct reports explicitly mentioning AI in their descriptions.

I don’t like to be critical of schools that release data about academic misconduct cases. But in this case, it appears the school didn’t release this information so much as it was asked, was obligated to answer a records request, and it did.

Note, these are integrity cases that only specifically listed AI or derivations thereof in their descriptions. So, it’s not a full view. Nonetheless, taking the largest number here, 66, that’s a non-credible number. The University of Florida has an enrollment in excess of 61,000, ranking it among the largest public schools in the country. If the numbers are accurate, last spring, just .1% of students were implicated in cases of AI-related academic misconduct.

In all seriousness, why bother?

There are, to paraphrase the Bible, none so blind as those who will not see. Something to keep in mind the next time you come across someone with a recent degree from Florida. Or Texas (see Issue 296).

Naturally, this debacle would not be complete without this, from the coverage:

University officials say they are attempting to teach students and faculty how to use AI responsibly, rather than banning the technology from the classroom.

I won’t even ask how that’s going.

And:

“We recommend that we don’t ban AI for everything but to carefully figure out where can we use AI with the students and where should they not,” [a school administrator] said.

But, I ask rhetorically, what is the point in creating places where students should not use AI if you’re not going to do anything when they do?

The same administrator said:

Accurate AI detection is not truly feasible, he said, and even the best AI detectors have a 4% false positive rate.
“This means that if you use them and rely on them, you’re going to send 4% of your students to the Dean of Students [Office for] Honor Code violation while they did not use AI,” he said.

That administrator is Hans van Oostrom, the director of UF’s Artificial Intelligence Academic Initiative Center.

He is misinformed.

The “best” detectors have false positives well below 1%. I can’t defend why van Oostrom does not know this. My guess is that he read a news article three years ago and considered the issue decided, even though the state of the science has shifted dramatically since then. No inquiry. No curiosity. No learning. He heard an answer and repeated it. Which, if you think about it, is very on brand for an “Artificial Intelligence Academic Initiative Center.”

But there is hope. At least someone at Florida is using their brain:

Shu-Jen Huang, a UF professor of mathematics, teaches a large mathematics class in which students complete assignments based on the computing platform MATLAB. At the start of the course, the AI class policy is made clear to students: They may not use AI to complete homework.
However, she said of 600 students, at least 80 were found using AI to complete their first assignment.
“Some students just do the minimum work. They just put everything to AI, and they get the results,” she said. “They are not learning anything.”

Eighty. Out of 600. On just one assignment. Fascinating. And no, they’re not learning anything, at least not about math.

The professor goes on to say how she made it difficult to get away with using AI on the assignments, but that her students “quickly found ways to work around these methods.” She adds:

AI creates an unfair advantage, she said, because some students spend hours trying to complete an assignment, while others finish in a matter of minutes.
“If you don’t want to spend that time, then I don’t think you deserve to get an A,” she said.

The article continues with a review of the process of misconduct cases at the school, including views from two students.

I’ll close by disclosing that I’m a Gator — a member of the alumni association and occasional donor (journalism and liberal arts). And I’m embarrassed. To the point of nausea.

Subscribe now

News Coverage from Australia

It was an important development in past weeks that a court and regulators in Australia had hit cheating provider Chegg with a six-figure fine for facilitating cheating (see Issue 428).

This week, one of the major papers in Australia ran a follow-up article.

I cannot read the whole thing because my subscription expired, and I’m not often motivated enough to pay for a full month or more of subscription fees for one article I will use for five minutes. But, from what I can read, here is the second paragraph:

But while the decision [to fine Chegg], handed down in March, puts academic integrity at the forefront of higher education, there is growing evidence students’ use of contract cheating companies to complete assessments has been replaced by reliance on AI.

Yes, students are using AI to cheat now. Nearly universally. And it has all but killed fee-for-service cheaters such as Chegg. That’s been covered frequently.

But that reality may spark one to ask — does this mean AI companies are the new billion-dollar academic cheating profiteers? Are they killing the Cheggs because they are just replacing them? Is OpenAI the new Chegg?

To me, the answer is clear — yes. And there are many reasons to think so.

Should Australia — the only country that actually seems to care about academic fraud and academic quality — bring regulatory or legal action against OpenAI, Anthropic, Gemini, and the others? Yes, they should.

Will they? Probably not. But we will see. Australian leaders have both fooled and impressed me before. Meanwhile, the rest of the globe does next to nothing to even pretend that the education credentials its institutions are awarding represent actual learning.

That’s where we are. We’ve been here for some time.

Subscribe now

A Personal Note

As many of you know, I care deeply about human writing and storytelling.

Yesterday, I was honored to be on NPR’s national news program 1A to discuss AI use in writing, publishing, and the need for disclosure.

I was joined by New York Times reporter Alexandra Alter, who covers media and publishing, and by Andrea Bartz, the best-selling author. If her name rings a bell, she sued the AI company Anthropic for using pirated copies of books to train its model, recovering a $1.5 billion settlement for authors.

If you’d like to hear the show, here is the link: https://the1a.org/segments/what-ai-authored-books-mean-for-the-publishing-industry/

Subscribe now

(429) New Research: Humans Not Great at Detecting AI Content. The Machines Are.

Derek Newton — Tue, 14 Apr 2026 16:37:46 GMT

Subscribe below to join 5,230 (+11) other smart people who get “The Cheat Sheet.” New Issues nearly every Tuesday and Thursday.

As always, The Cheat Sheet is verified human-written:

New Research: People Not Good at Finding AI-Created Content, AI Detection is Much Better

Stop me if you’ve heard this one before — people are not good at spotting text created by AI, the AI detectors are better.

You may have heard that before because we’ve covered similar research several times (see Issue 253 and Issue 332). Those are different than the studies showing that AI detection works, of which there are many. But those two, like this new one, show the comparison between human detection rates and those of the machines.

[Note: one paper found that a majority vote of five people who had significant experience reading for AI could match detection rates of the technology — but that’s different. And wildly impractical. See Issue 357.]

This new paper, published in Nature last month, is by:

Matheel AL-Rawas, Omar Abdul Jabbar Abdul Qader, Galvin Sim Siang Lin, Yew Hin Beh, Muhammad Annurdin Sabarudin, Yee Ang, Jun Fay Low, Johari Yap Abdullah, and Tahir Yusuf Noorani.

Most of the authors are affiliated with dental schools in Malaysia. Qader is based in Iraq; Abdullah is affiliated with a school in India.

The team used young scholars, presumed grad students, and AI detectors to evaluate 150 abstracts for research papers in dentistry. Seventy-five were human-written, 75 were GPT generated. And though they used research papers in a specific scientific field, the team was focused on academic integrity implications.

Here’s the key finding, from the abstract:

These findings suggest that relying on human judgment alone is insufficient for identifying AI-assisted academic text and that selected detection tools may support academic integrity safeguards as AI writing technologies continue to evolve.

Of note, the team tested four AI detectors:

GPT-2 Output Detector, Writefull GPT Detector, GPTZero, and Turnitin similarity detection

The first two are so bad, one does not even exist anymore — shuttered by its creator. GPTZero and Turnitin, on the other hand, are in wide use and credible. It’s not clear to me whether the Turnitin “similarity detection” is their AI-detection product. It’s noted differently, but the results lead me to believe it must be. Either way, the paper authors do us a great service by testing it and GPTZero.

It’s also important to note that the human reviewers were told specifically to look for AI-created text. Human reviewers were also instructed not to use AI-detection tools or the Internet to help evaluate the abstracts.

Here are some highlights from the paper, namely the key one:

Overall, young academicians, regardless of institutional category, had difficulty identifying the origin of AI-generated abstracts, whereas GPTZero showed the highest discrimination accuracy (90.0%).

The overall accuracy of human reviewers in detecting abstract origin varied from 44% to 76%

And:

GPTZero was the most accurate AI output detector, with an accuracy rate of 90%. This was better than GPT2 Output (55%) and Writefull GPT (58%). The similarity detector was 94% accurate

Again, I am confused as to why the similarity detector was counted separately and, for that matter, what it is checking. Either way, 90% is better.

This is also interesting. Citing a different study in the dental sciences, the paper says:

Researchers conducted a study evaluating the capacity of novice teachers (average age 25) and experienced teachers (average age 42) to identify AI-generated content in student submissions. The findings revealed that both novice and experienced teachers encountered difficulties in differentiating between AI-generated and human-authored texts

The team also writes:

Generally, research has shown that humans have difficulty in differentiating between text produced by AI and text generated by humans. Automated systems, specifically AI bot detection, excel at distinguishing content. AI-generated composition often lacks specificity, creativity, tends to overstate particular instances, and exhibits a distinct writing style characterized by the use of predominantly predicted words, in contrast to human writing. AI systems such as GPTZero have achieved notable success in probabilistically detecting AI-generated text. In this study, the GPTZero detector was very effective in differentiating AI abstracts from the original abstracts compared to the other two detectors, and in our observation, there were more false outcomes by the GPT-2 Output and Writefull GPT detectors, scoring 48 and 35 AI abstracts, respectively, out of 75 as abstracts with low fake content

Even more gushing over GPTZero:

In our opinion, GPTZero is a very useful tool for the academicians to identify the source of the text or abstracts.

And more:

If GPTZero demonstrates high accuracy, universities may implement it as a standard tool for detecting AI-generated submissions, thereby enhancing academic integrity policies.

The paper also deserves credit for saying — as so many have said so many times:

institutions should refrain from excessive dependence on this tool and incorporate human verification in conjunction with AI detection mechanisms. AI detectors should only be used as decision-support instruments …

Agreed. I do not think anyone disagrees.

And, though not directly on point to academic misconduct, here is something to keep in mind about teaching students to use AI because, we are told, they will need to use it in their careers:

Several conferences and journals have already established policies that explicitly forbid the use of ChatGPT in their research output. In the midst of publishers' efforts to revise their policies, it is the responsibility of all human authors to understand that any violation of these policies will be considered scientific misconduct, similar to plagiarism of existing research

Can confirm. Most academic journals prohibit using AI-created material.

Finally, not to underscore this too much, the paper found:

While GPTZero and Turnitin similarity detectors exceled in accurately classifying abstract categories, young academicians from either research and non-research based governmental or private universities struggled in the identification of HI and AI-generated abstracts

One note on this paper, from me, is that the source for the “human authored” papers used in the study does not seem entirely reliable:

A total of 75 titles and original abstracts were collected using random sampling from recent issues (published in the first five months of 2023) of 8 high-impact dental journals using systematic sampling.

Since we know that research teams are using AI to write, or at least refine, their papers at a significant clip, there’s no reason to assume that 75 articles from 2023 are free of AI-created text.

Something to consider.

Anyway — people, not great at this. The good machines — much better.

Subscribe now

From Turnitin, On “Learning Integrity”

In Issue 426, I wrote about some surprising comments from Turnitin, the leader in plagiarism and AI detection.

Below is a reply, or additional information from Chris Caren, CEO of Turnitin, unedited:

AI is disrupting institutional integrity policies, and there should be healthy debate. In most cases, we are seeing the “acceptable use of AI” defined at the individual educator level, and the definition can also change by assignment. From our view, the evolution to learning integrity is an expansion of existing academic integrity frameworks, because analysis of the final product is no longer enough; proof of a high integrity learning process (in this case writing) is also needed.

When we reference rigid institutional frameworks, that is speaking to black and white integrity policies that work when it comes to “traditional” plagiarism but require more complexity when it comes to AI use. For example, AI use is allowed in some educational contexts, and not others, while traditional plagiarism is never allowed.

AI can be used to “polish” the writing of a student, which is acceptable to some educators and not to others. Educators also disagree on whether using AI to create an essay outline for a student to write to is acceptable. This is why I feel strongly that AI detection needs to be paired with proof of process to be a lasting solution in education.

Responsible use of AI in education is case-dependent. Thus, we make the case for “learning integrity” that includes bans on AI in some cases and an exploration of how and why AI was used in others. There is no one-size-fits all when it comes to responsible use of AI in education, and we want to partner with institutions as they navigate their definitions.

Subscribe now

Two Stories on Cheating, From India

In India, cheating is at an entirely different level.

Here are two relatively recent stories from India to make that point. I’ll just put in the headlines and links. Enjoy:

Student Assaults Invigilator After Being Caught Cheating

And:

81 Teachers Suspended for ChatGPT Cheating

Yes, teachers.

Subscribe now

(428) Misleading About the Adelphi/AI Detection Case

Derek Newton — Tue, 07 Apr 2026 16:20:01 GMT

Subscribe below to join 5,219 (+12) other smart people who get “The Cheat Sheet.” New Issues nearly every Tuesday and Thursday.

As always, The Cheat Sheet is verified human-written:

Needed Detail About the Adelphi, AI Detection Case

Prologue

At the recent conference of the ICAI, the International Center for Academic Integrity, I heard a presentation from Trinka, one of the conference sponsors — a lunch talk to the entire conference. Trinka is an AI writing platform, it sells AI writing tools, offering to students that they can “write and rewrite with AI.” Trinka also has a grammar checker. And a paraphraser. And an AI detection tool.

The suite of services feels much like cheating provider Grammarly.

I have no idea what an AI writing company was doing on stage at a conference about integrity. That is deeply troubling, but not my point.

Much of their presentation was about how AI detection did not work and was bad, which is entirely predictable because you can’t sell AI creation if it can be easily spotted. Every AI provider from OpenAI, to Grammarly aggressively hawks the idea that AI detection does not work or is dangerous. Mind you, Grammarly and Trinka both have their own AI detection tools.

Nonetheless, this particular presentation went way too far, in my view — veering from sales tactics to outright deception.

Trinka’s representative told attendees that using AI detectors may ruin students’ lives and that schools could be sued for using them. To make this point, the speaker highlighted a legal case regarding a student at Adelphi University who successfully sued to overturn a finding of academic misconduct for AI use, strongly suggesting that this outcome was because the AI detection system used at the school — Turnitin — gave a false positive. Therefore, the company argued that AI detection was unreliable and dangerous and the best course was (surprise!) to let students use AI while surveilling their work progress.

The Case

Hearing the presentation, I was embarrassed that I had not heard about this case and made a note to look into it. And I did.

It is true that a student at Adelphi was found responsible for academic misconduct for having (allegedly) used AI to write a paper. Turnitin scored the paper as 100% AI-created. Throughout the process, the student denied using AI. After being found responsible, the student appealed the finding and eventually sued, seeking a reversal of the outcome. A judge vacated the finding of responsibility while also clearing the student’s record.

But here is the thing Trinka conveniently left out — the decision had nothing to do with the AI score or the AI system used. There was no finding that Turnitin was not correct. No finding that the student did not, in fact, use AI as accused.

Instead, the judge ruled for the student and against Adelphi almost entirely because the hearing and appeal process was flawed. Most notably, the professor who initiated the integrity case cast doubt on the objectivity of the hearing process. At one point, the professor, in an e-mail to the accused student, urged the student to appeal, writing:

it seems like their system defaults you as responsible unless you contest the claim again

When the professor who initiated the case encourages the student to appeal the outcome and questions a default judgement, that’s a problem. And the Judge noticed.

Further, despite policies requiring it, the student was denied access to an advisor or counselor for the hearing. He was also not allowed to present evidence in his defense, evidence that included 0% scores from other AI detectors, Grammarly and zeroGPT. Yet further, the Judge noted that the appeal of the finding of responsibility was made to the exact same person who made the initial determination — an appeal that wasn’t credible in other words.

For the record, I have serious doubts about the accuracy of scores from Grammarly and zeroGPT while generally trusting Turnitin’s scoring. I have zero first-hand knowledge about this case specifically, but I nonetheless find it very unlikely that Turnitin would score a document at 100% when it was really 0%. But more on point, the Adelphi case in no way relied on this question anyway.

The case went the way it did because Adelphi screwed up their process. They ignored their own policies and created an appeal that was a joke. They afforded the student no reasonable opportunity to present his case, even though he was assured he could. All to the point where, to repeat, even the professor who suspected cheating questioned it.

Consider, from the Judge’s order (legal citations removed):

Moreover, the [Code of Student Conduct] requires that a student be provided with a "meaningful opportunity to be heard" and that any findings of misconduct be based upon a "preponderance . . . of the relevant information." However, in issuing the Denial, [the administrator] failed to even consider the Petitioner's evidence in the form of the two AI detections programs, which indicated that the Essay was "human written." Finally, as previously noted by this Court during oral argument, the mechanism employed by Adelphi in adjudicating the Petitioner's alleged misconduct functions such that the individual who issued the Determination - [the administrator] - was the identical person charged with the responsibility of entertaining the appeal.However, such policy would deliberately thwart a student's right to an avenue of meaningful "appeal" as expressly provided in the [code of conduct] and renders the term inconsequential

And noting that the school’s policies say that:

"[a]ll students have the right to . . . [b]e accompanied by an advisor of choice who may assist and advise a reporting individual, accused or respondent throughout the University disciplinary proceedings including all meetings and hearings related to such proceedings"

The Judge wrote:

the Petitioner was not afforded the opportunity to confer with an advisor of his choice regarding the Violation

Yikes.

The Outcome

Nearly exactly as we saw play out in a legal challenge to remote exam proctoring in Ohio (see Issue SE2), a student won a legal case that had nothing — or nearly nothing — to do with the accuracy, reliability, or appropriateness of the technology. In both cases, the schools screwed up their process. In Ohio, the school messed up delivering the technology. In New York, where Adelphi is, the school messed up the hearing, adjudication, and appeal process.

Also in both cases, people who wanted to attack the assessment technologies shamelessly and unethically misrepresented the outcomes. Room scans during remote proctoring are not, and never were, unconstitutional. Adelphi was not sued because they used AI detection technology, and the school certainly did not lose the case because of it either.

I don’t have much more to say other than that I wish we could stop misleading teaching and administrative professionals, scaring them with misinformation just to sell a product. Yes, I know, I expect too much.

Subscribe now

Chegg Fined $500,000 AUD for “Contravening Academic Cheating Laws”

Oops.

Chegg, which has laughably maintained for years that it was not a cheating provider, has been fined a nice $500,000 AUD, about $345,000 USD, for “contravening academic cheating laws,” according to Australia’s higher education regulator.

Here is the first paragraph of the announcement:

The Tertiary Education Quality and Standards Agency (TEQSA) welcomes today’s decision of the Federal Court of Australia in proceedings brought by TEQSA against Chegg Inc. The Court found that Chegg contravened academic cheating laws on three occasions through its ‘Expert Q&A’ service, and ordered it to pay penalties of $500,000 and costs of $150,000. This decision marks a significant outcome for the integrity of higher education sector and is an Australian-first for contraventions of Australia’s academic cheating laws.

So, my bad — $650,000 AUD, or about $448,000 USD.

I’ll highlight here that the fines and costs were for three violations. Three. I’m willing to bet that there are closer to 30,000. Nonetheless, the goal was the message, and I suspect it was delivered.

Class Notes

I apologize that The Cheat Sheet has missed two normal publication dates. A week ago, I was on my way back from Arizona, from the really amazing ProctorioX conference, where I was honored to be presenting. Then last Thursday, I was waiting on a file to add here, which did not come in early enough. Sorry.

In any case, should anyone be interested, here is a link to the slides from my presentation at ProctorioX — Counting Down the Top Three Threats to Academic Integrity.

Subscribe now

(427) University Reports 95% of Its PhD Research Contains AI Plagiarism

Derek Newton — Thu, 26 Mar 2026 17:34:11 GMT

Subscribe below to join 5,207 (+9) other smart people who get “The Cheat Sheet.” New Issues nearly every Tuesday and Thursday.

As always, The Cheat Sheet is verified human-written:

University in India: 95% of PhD Theses Contain AI Plagiarism

According to local coverage, more than 95% of recent PhD Theses at Lucknow University (University of Lucknow, enrollment of about 20,000) contain suspected plagiarism.

From the coverage:

Data accessed from the Tagore library shows of 121 theses submitted, 116 carried similarities to previous research, pointing out that AI-generated content was used to write them.

A company called DrillBit was used to do that AI detection, the paper reports. Also according to the article, the school allows up to 5% AI in papers, about 180 words, but these examples were in the 25-50% range, the school says.

Also, according to the reporting, papers flagged for AI overuse or misuse are returned to the student for correction. “To make it original,” the paper reports.

This is, quite clearly, silly. If the goal is to limit AI use in what is supposed to be original research, simply giving students a do-over is a waste of time. Tell me what the disincentive is.

You’re a student, use AI all you want. The school will circle the parts you need to change, and you can edit — probably with a humanizer, the skeptic in me wants to say. I’ll leave it to others to judge whether that’s quality scholarship or teaching good scholarly habits. But I think the fact that 95% of papers have AI is all the answer you need.

Of note, the story also says that the school will soon be switching to Turnitin’s software to check more assignments at different levels. I guess that’s good. But if the process is the same, I’m not sure there’s much point.

Subscribe now

Looking at Auto-Typers, from This Isn’t Fine

Our friend and colleague Joseph Thibault over at the outstanding newsletter This Isn’t Fine, has an article out this morning on auto-typers. It’s well worth your time to review.

It’s a test of more than a dozen auto-typers, the software designed to mimic human typing with the goal of cheating process trackers such as Google Docs revision history — have AI do the writing, then have an auto-typer peck it out in Google Docs and (presto!) you have a document with “proof” that it was written by hand.

With these, student work could be flagged by a good AI detector as AI, but the student protests and offers their completely manufactured Google Doc revision history as evidence. Without knowing, or at least being able to suspect, that the Google Doc data is bogus, schools would likely dismiss misconduct cases which follow this pattern. Many have, in fact.

In addition to students skirting accountability and cheating their institution, classmates, and the public, a pattern along these lines looks as though the AI detector was bad when it wasn’t.

Quoting This Isn’t Fine:

In my opinion, humanizers are to AI detection, as auto-typers are to process tracking.

That’s right. Only they’re actually worse. Humanizers have a plausible reason to exist — thin, but plausible. It’s possible that someone may wish to see how their writing feels when it’s been changed by AI. Or, writing in a secondary language may feel more confident if AI “humanizes” their work. Thin — but possible, I suppose.

For auto-typers, the only raison d'etre is to facilitate fraud. They exist only to create evidence of effort — not to validate it, but to create it. As these tools grow in reach and grip, relying on Google Doc history, or other kinds of process trackers, is foolish.

In any case, if you’re involved in or care about the process of academic misconduct, jump over and read the article.

Subscribe now

Reader Submission: Doomed for Divorce

Space in The Cheat Sheet is always open. Below is a submission we received from Ian McCullough, whose somewhat shortened bio includes that he is:

an EdTech marketing executive with 25 years of experience spanning consumer products, institutional sales, and corporate learning. He has held leadership roles at 3P Learning, LeapFrog, Pearson Foundation, and Electronic Arts. At Turnitin, he originated and scaled the K-12 regional marketing function across four progressive roles over six years, contributing to double-digit ARR growth for the business unit and forging strategic partnerships with NEA, AFT, and NASSP.

Here is his commentary, unedited:

Instruction and Assessment Are Doomed to Divorce
Both functions are under crippling pressure pushing in opposite directions. Academic institutions have no choice but to split them up.
The chorus proclaiming the death of the resume grows louder every day. Every administrator in American Higher Education should be losing sleep over this. While debates about whether the value of formal education is humanistic or economic are eternal, many people do indeed pursue college degrees for career advancement. Against that reality, a fiscal truth becomes clear.
If the resume is dead, then so is the degree. If colleges and universities want resurrection, then a marriage must end. Instructors should no longer assess their own students.
The inevitable divorce of instruction and assessment was set in motion decades ago. While the institutional template of universities as temples of scholarship and research dates back almost a millennium, the industrial revolution sparked demand for businesspeople and engineers. Universities developed new degree programs and became avenues of socioeconomic advancement. After World War II, the G.I.Bill provided tuition subsidies and later laws introduced loan guarantees. When government-backed debt became the fuel of socioeconomic advancement for the masses, competing incentives collided. A well-researched cycle began.
Subsidizing tuition led more people to enter degree programs.
Applicants to programs surged amid limited capacity and prices rose. The government responded with increased grants and loan limits
More money led institutions to seek more applicants. Upgraded facilities and amenities made campus life more attractive… and the cycle continued.
Over time, the default model of undergraduate education became customer service. The death of the resume, however, provokes a new question: if students are customers, then what are they even buying anymore?
Instructor success in undergraduate courses is generally measured by popularity. Many instructors are adjunct faculty with little-to-no job security. They face potent incentives to appease students by inflating grades and turning a blind eye to academic misconduct. When the quality or integrity of a student’s work is challenged, the institution risks lawsuits. Now doubts about the value of the Masters degree are mounting.
For employers to value academic credentials, the divorce of instruction and assessment needs to happen. Arguably, it already has.
Resumes were already portfolio statements and job interviews were already oral examinations. AI tools can now ingest work samples and interview responses and make instant recommendations about what candidates to bring in for paid tryout periods. Since employers are the ones with the strongest incentive to get assessment right, they are now the ones who are doing the assessing that matters.
Especially with the new Federal loan institutional accountability requirements, the question now is how Higher Education will adapt.

Subscribe now

(426) Two Webinars Next Week

Derek Newton — Thu, 19 Mar 2026 17:43:05 GMT

Subscribe below to join 5,198 (+45) other smart people who get “The Cheat Sheet.” New Issues nearly every Tuesday and Thursday.

As always, The Cheat Sheet is verified human-written:

Upcoming Learning Opportunities

There are two upcoming opportunities to learn more about academic integrity.

Fraud Triangle

One is a webinar on Wednesday, March 25, with Joseph Thibault and Dr. Mark Glynn. Details, provided by the hosts, include:

Pressure.
Opportunity.
Rationalization.

Three things that impact our decisions for how we do things. They were popularized in the 1950s in thinking about fraud...

In this webinar, we’re going to leverage the framing to think about minimizing each leg of the fraud triangle during assignment redesign to promote assessment validity. (Academic fraud is the most extreme end of academic dishonesty.)

Dr Mark Glynn and Joseph Thibault will team up to discuss the “fraud triangle” during a free webinar on March 25th at 8am. We’ll discuss many things including:
-Cressey’s original model
-Reason’s Swiss Cheese Model
-Practical application of the triangle by critiquing real assessments
-1st hand research from Mark and DCU, and
-Ways that you can minimize opportunity, pressure, and rationalization for the work and assessments you value in the classroom.

This is the third installment of our series on AI and Assessment featuring Dr. Mark Glynn to discuss the fraud triangle, academic integrity, and how you update courses, assessments, and policies to guide students to and through authentic work.

To join, use this link.

Challenges and Developments

A day before that one, this time on Tuesday, March 24, I’ll be on a panel with the title:

The State of Academic Integrity: Challenges and Developments in the Education Sector

Here is the summary:

Academic integrity is facing significant challenges due to artificial intelligence, contract cheating, digital assessments, and changing learning behaviors. Institutions must adapt by ensuring compliance with data protection and security standards. Key strategies include updating policies, adopting secure assessment technologies, enhancing faculty training, and promoting a culture of academic honesty for lasting success.

The second part of the session will be:

Part 2: A Practical Case Study – Secure and Compliant Implementation
In the second part, Cathleen Mahoney, CEO of Bright Language, a language proficiency assessment partner of a large number of organizations worldwide, will share practical experiences with implementing online proctoring.

To join this one, use this link.

Subscribe now

Turnitin Dismisses “The old ‘deter and detect’ mindset of academic integrity”

I admit, this really surprised me.

Turnitin, the market leader in academic integrity solutions for decades and the OG on which the framework of integrity technology has been built, has, it seems, kicked all of that to the curb.

In a statement in their corporate blog a few days ago, the CEO of the billion dollar company opened with:

The old “deter and detect” mindset of academic integrity is no longer scalable or effective when compared to the process-based approach of learning integrity.

I get what he’s trying to say. In fact, he says it in the next sentence:

We have to shift to assessing the learning process and not just the final product to understand how and why AI was used.

I agree, process is important. But, in fairness, it always has been. That’s not new. And, of course, Turnitin has a new(ish) product that will track and evaluate the process of learning, what they’re terming learning integrity. This product, like others from other companies, puts eyes on the development of academic work — the process — and not just the end artifact.

No issues with any of that. I think it’s good practice.

Nonetheless, it’s jarring to hear Turnitin declare so clearly that the “deter and detect mindset” of academic integrity is no longer, essentially, viable. This feels like a titanic shift in philosophy and it’s deeply troubling. Turnitin built and sold the original, scalable plagiarism detector to deter and detect copy-and-paste academic work. But that approach — Turnitin uses “mindset” — is no longer effective, the company says.

Seriously, that reads to me like surrender.

Lest you think I am being too strong, the third sentence in this message from the CEO includes this:

a gap has emerged between rigid institutional policies and classroom reality.

Turnitin, as I read this, thinks that “rigid institutional policies” do not match with, or work in, reality.

Wow.

The message is clear. Turnitin says ‘detect and deter’ is old thinking and not effective. They say we need to “shift,” and that “rigid institutional policies” are not reality.

Wow. Again.

Let me tell you how I read that. I read: Students are going to use AI, no matter what we tell them or what we do. We cannot, or simply don’t want to, police or deter that anymore. Let ‘em go. Let’s watch how they use AI instead.

I am sure some readers of that will agree. I don’t. And I am still surprised to hear this from, of all places, Turnitin.

In fairness, this message from the CEO is muddled and, I will argue, contradictory or inconsistent. Further down it includes, for example:

If a student "outsources" their thinking to AI, they aren't just breaking a rule, they are not becoming employable -- and slowly destroying the value of a degree from their school.

True. But rigid policies against this are not reality, they are old mindsets. Square that circle, go ahead.

There’s also:

In a high-stakes certification (like medicine or law), "zero-AI" guardrails remain vital for public trust.

Another square that does not seem to circle with the idea that rigid policies and AI deterrence and detection are ineffective.

According to Turnitin, we can, and should, ban AI when the stakes are high. That’s not an old mindset, that’s for trust. Which is something we only need in places “like medicine or law.” It’s not that banning AI is bad, says Turnitin. In fact, it can be “vital,” they say. But it’s still somehow not “classroom reality.”

I confess, I don’t get it. I don’t get what Turnitin is trying to do here. Or why.

But I do firmly think it’s very unhelpful in what I view as an existential struggle for the very value of learning and the bright lines of integrity. And those, to me at least, are related — not outdated.

Subscribe now

Class Notes

It’s been two weeks since the last Issue of The Cheat Sheet. Sorry. For the past ten days or so, I’ve been down with a serious respiratory virus. Whatever this is, it sucks.
It was my pleasure to run out to Denver two weeks ago to present at the ICAI conference, the International Center for Academic Integrity. I’ll see if I can get the presentation I did with the amazing Joseph Thibault. Spoiler alert: Essay Mills — They’re Back.
Also, I don’t understand why there aren’t a thousand schools represented at ICAI. For symbolism alone, ICAI ought to be one of the largest academic gatherings every year. To me, it says a ton when schools don’t show up to a conference about integrity.
Speaking of ICAI, if you missed Pangram’s presentation on the state of AI detection, you are behind. I’ll see if I can find it and share it. But one more point on this: Not 30 minutes after hearing the best presentation on why AI detection is possible, how it works, and how it’s getting better every day, an attendee at ICAI in a different session started a post-presentation question by saying, “Since we all know that AI detection does not work…” At ICAI. I just can’t.
And remember that $75 million jury verdict against Course Hero (see Issue 425)? I did a little piece on it in my tiny corner of education coverage.
Subscribe now

(425) Jury: Course Hero/Learneo Violated University Copyrights, University Awarded $75 Million

Derek Newton — Thu, 05 Mar 2026 18:48:04 GMT

Subscribe below to join 5,153 (+49) other smart people who get “The Cheat Sheet.” New Issues nearly every Tuesday and Thursday.

As always, The Cheat Sheet is verified to be human-written:

Course Hero/Learneo Violated Copyrights of Post University, Jury Finds

Before we start, Learneo is the company that is Course Hero. Rebranded in late 2022. Learneo is a portfolio of pernicious products including Course Hero, Quillbot, Scribbr, and Cliff’s Notes. These products are the antithesis of education. In this article, I’ll use Learneo. But they’re the same.

The news is big — a jury in federal court has found that Learneo violated the copyrights of Post University, that’s according to a note I received last night from the Post team.

For background on the legal challenge, see Issue 252.

Here is most of the e-mail I received:

Post University has prevailed at trial in its lawsuit against Learneo related to their Course Hero website, following a federal jury’s consideration of the evidence and arguments presented. The jury found in favor of Post University on its claims related to the unauthorized use of copyrighted academic materials, marking a significant outcome in the case.
“Today’s verdict sends a clear message: educational content cannot be taken, altered, and sold without permission,” said Kimber Summers, Chief Regulatory Officer and General Counsel. “This outcome supports academic integrity, and we are grateful that the jury recognized that academic materials deserve legal protection.”
This verdict concludes a closely watched case addressing the treatment of copyrighted educational materials in the digital marketplace and reinforces the role of the courts in setting clear boundaries for commercial content platforms.

This morning, I received this note as well:

Late yesterday, a federal jury in Connecticut awarded $75 million to Post University in its closely-watched Digital Millennium Copyright Act (DMCA) case against the academic file-sharing site Course Hero. The jury found more than 3,000 DMCA violations and awarded Post University the maximum penalty after deliberating for nearly 14 hours.
At issue were thousands of Post University tests, study guides and other documents uploaded to the Course Hero website with copyright management information removed and replaced with Course Hero’s own identifiers. The jury found this violated the DMCA’s protections against tampering with rights-management information.
The large verdict will likely set a precedent for similar cases involving uploaded materials, in both the academic sphere and beyond.
Harris St. Laurent & Wechsler LLP represented Post University in the 5-year case, along with co-counsel from Dilworth IP LLC.

As I have said before, Learneo has been profiting from stolen academic material for years — selling it to students to help them cheat. My opinion. But maybe this outcome lends a little more credibility to that view.

Anyway, it’s big for Post University because, although it’s not specified in the information I have, I would expect that Post’s academic materials will be removed from Learneo’s platforms, if they’re not already. This will make cheating at Post more difficult and improve the school’s learning outcomes as well as its reputation.

Post also cashed in — $75 million is real money. Every university in the country would love to have it.

And as big as it is for Post, this win is bigger in general.

For one, Post is not nearly the only school to have academic materials used and sold by Learneo without permission. Every school, training provider, assessment system has been compromised by Learneo — for years. Course Hero, right now, shows 73,183 documents from the University of Florida, for example. I’d wager that Learneo has the rights to — give or take — none of those.

Continuing, this could cripple Learneo. They have a ton of money, sure. Pirates often do. But not many companies can write a check with eight digits in the dollars place. And if other schools smell blood here and file similar cases, well — that would be great for integrity, very bad for Learneo.

I left in the name of the law firms, should anyone out there be interested in speaking with them. Cough.

Another win here is that this verdict probably allowed Post University, and may allow others, behind Learneo’s walls, getting them access to see what the company really has, how it hides it, and how it sells it. We may even learn what Learneo knows about who its customers are and what they’re really doing. I mean we all know. Learneo has just been allowed to play dumb about it. Maybe that ends.

Finally, there’s also pure symbolism. And that cannot be bigger. As cheating giant Chegg circles the drain (stock is down to just 64 cents as of this morning), Learneo having a humiliating defeat in court is beautiful.

To get more about why it’s beautiful, scan Issue 252. Or use the search feature here for “Course Hero.”

Subscribe now

Oberlin College Cancels Contract with Grammarly

According to local coverage, Oberlin College has “discontinued” its contract with Grammarly over concerns the tool can be used as a cheating agent.

To be clear, it is a cheating agent. No disclaimer needed.

From the coverage:

“Grammarly is no longer just a grammar checker,” Andrea Simakis, director of media relations, wrote in an email to the Review. “It now includes tools that generate ideas, suggest sources, predict grades, and ‘humanize’ AI-written text. Faculty felt those features blur the line between editing and doing the intellectual work for students, which conflicts with the learning goals of Oberlin’s writing requirement.”

Absolutely right. Grammarly is an AI text generator and has expansive tools to help users avoid detection.

I love that, according to the reporting, the decision to ditch Grammarly came from:

visiting experts in academic integrity and AI, who came as part of Oberlin’s Year of AI Exploration

I’m blown away that a school had events about AI and invited integrity experts. Absolutely amazing. I’m every bit as surprised that the school took action based on the advice of experts. It’s so logical, so sound that I struggle to believe it.

I’m sharing another quote from the article so as to underscore the rare observed reality happening at Oberlin:

“Utilizing generative AI at any stage of the writing process, which now includes some of Grammarly’s tools, is antithetical to our mission and values,” Writing Associates Program Coordinator Kayla Kim, OC ’25, wrote in an email to the Review.

Early frontrunner for Quote of the Year.

It’s also a downright delight to see a school unplugging a cheating provider for once, instead of unplugging the tools that can help deter and detect cheating — inexplicably keeping Grammarly, but dropping Turnitin, for example. That’s backwards.

Oberlin is right. And I really do suspect and expect that it won’t be the last school to cut ties with Grammarly. It should not be.

Subscribe now

Sal Khan Doesn’t Know What He’s Talking About

Fine, that’s a salacious headline. I’ll own that. But at least as it applies to this article, it’s true.

The piece is basically Sal Kahn, of Kahn Academy fame, giving “his five priorities tomorrow’s schools should view as non-negotiables to confront the unprecedented AI Age.” It’s a tired trope at this point and it’s embarrassing, in my view, that media keep talking to him.

Nonetheless, I was really impressed to see that one of Kahn’s priorities was:

Priority 2: Acknowledge Rampant Cheating
Research from the Center for Democracy & Technology suggests cheating through generative AI rose from 58% to 70% in the 2023-2024 school year. Even this hefty statistic may be understating the problem’s ubiquity. That’s because many of the AI-detection programs don’t work. “These tools that claim to detect AI are snake oil,” says Khan. He pointed to high false-positive rates while also noting that so-called humanizer tools “can make AI-written work look human anyway.”

A few things.

Again, it’s great to see the gross realities of cheating make any list at all. And I’ll say that the 70% number is low. But whatever. It’s a problem and Kahn is right to say so.

But the rates of cheating aren’t being set by AI detection software. I don’t know where this idea came from. In fact, in the link the writer cited, the 58% to 70% is a survey of high school students. And that survey is not about cheating at all. Here’s what it says:

the percentage of teachers who reported using generative AI rose from 51% to 67% while high school student use increased from 58% to 70%.

Using AI — not cheating. So, that’s a total, total miss. And, again, even if those stats were about cheating, we cannot assume that the rates are actually higher because AI detection may be unreliable. That’s just odd. I mean, invented. And odd.

But this is about Kahn. His quote about AI detection and snake oil is wrong. He has — as I put in this headline — no idea what he is talking about. None. Awfully confident. Terribly wrong.

His comments about humanizers — also wrong. Making writing seem more human (to fool AI detection) is what they claim to do. That is true. But the good AI detection systems spot those like 100% of the time.

There, Kahn is so misinformed as to be embarrassing.

But if a priority is to acknowledge cheating — great. What do we do about it? The piece continues:

The solution isn’t to bury our heads in the sand, or worse, accept massive cheating as our new normal. Not if we care about our children. Instead, schools need to double down on priority one. Afterwards, assessments can determine if actual learning took place inside or outside the classroom. There’s also no reason not to bring back blue books and the rigor of in-person writing and oral exams to demonstrate genuine aptitude in real time.

I agree — the solution is not ignore it. It’s also no solution to accept massive cheating. Kahn says the answer is “double down on priority one” which is, by the way:

Priority 1: Lean into the Basics

Genius. That will fix our cheating crisis. The basics. Got it.

Why is anyone wasting our time and attention with this? For the record, to Kahn, the basics is “strong critical thinking and general knowledge.” Sure. And leaning into that fixes our cheating crisis how, exactly? Sorry, not just leaning in, but doubling down on leaning in — a lean in, lean in, I guess.

But, of course, once we lean in, assessments will work. Like pixie dust.

Also, you know, why not bring back blue books and in-person writing and oral exams? Sure. I mean, at least that does relate to cheating. And I am for all three of those things as tools to limit academic misconduct. But they are wildly impractical in most school settings and will do nothing to limit the use of AI and other self-defeating shortcuts on literally every other piece of academic work. Including “the basics,” no matter how hard you lean in. Or double down.

It’s useless word salad.

And here’s where I choose to remind readers of an interview Kahn gave in 2019 in which he said, “Now that I run a school, I see that some of the stuff is not as easy to accomplish compared to how it sounds theoretically.” [My source is Justin Reich’s outstanding book, Failure to Disrupt.]

To which I say — no kidding. Nothing he said in this most recent interview was worthwhile. And most of the specific things were just incorrect.

My only wish is that someone aside from me would fact check this junk and correct the record. OK, fine, I have another wish. I wish people would stop interviewing him.

Subscribe now

(424) LMS Killer Einstein Rebrands, In Hours, Then Dies

Derek Newton — Thu, 26 Feb 2026 19:11:26 GMT

Subscribe below to join 5,102 (+15) other smart people who get “The Cheat Sheet.” New Issues nearly every Tuesday and Thursday.

As always, The Cheat Sheet is verified to be human-written:

Meet the New Einstein. Gone: Brazen Cheating. For a Minute: Friendly Tutor. Now: Just Gone.

In our last issue, we introduced you to Einstein, the AI agent chatbot designed and sold to do all your academic work for you, literally it said, while you were asleep. It, and other agentic bots like it were going to kill learning management systems (LMS), I wrote.

Then, not too long after our Issue went out, a reader wrote in asking where I’d found the quotes from Einstein that were in that Issue. She could not find them. She asked, did I sign up and go further in? Or did Einstein remove those quotes and change, as they say, on a dime?

I checked.

Sure enough, gone was every single line I’d quoted — and I did not quote nearly all the ridiculous and shameless garbage which made it painfully clear that Einstein was there for only one reason, to cheat.

The Receipts of Change

For the record, there was no question I’d quoted Einstein’s page correctly. I literally copied and pasted the text into The Cheat Sheet. But here, thanks to The Wayback Machine, are the receipts, how Einstein looked two days ago when I wrote and sent the last Issue:

Clear as day — Einstein “logs into Canvas” and “submits your homework — automatically.”

It’s easier, I think, to compare these apples to apples. Here is the same section of the Einstein homepage, hours later:

Gone is logging into Canvas and doing your homework for you. Now Einstein is a “personal tutor” that “connects to your Canvas” and “guides you step by step.”

Here too is the section below that greeting, as it was on Tuesday morning:

On Tuesday morning, Einstein “logs into Canvas for you,” and “works while you sleep,” and “reads the assignment, solves it, and submits it directly.”

On Tuesday afternoon all that is deleted:

Now Einstein watches lectures “with you.” It can break down key concepts, create summaries. It now “runs through practice problems.” The scrub literally said, “Instead of giving you answers” it “walks you through the thinking process so you actually understand.”

Here also is that question I quoted from the FAQ on Tuesday morning:

In the Tuesday afternoon version, that’s simply erased.

And so I don’t bury part of the headline, as of this morning, Thursday morning, Einstein itself has been erased, it seems. When I went to the site to catalogue the difference, here is Einstein now (about 10am ET, 2/26/26):

RIP Einstein.

Tutorwashing

Even if Einstein stays dead, it won’t matter. The genie has been out of the bottle for a long time now. There are a hundred Einsteins. In 90 days, there will be a thousand. The LMS is still under existential threat. It will still be up to the LMS makers whether they decide to fix it or die. And whether the people who use and pay for LMS systems think that just addressing agentic chatbots is enough — blocking the Einsteins while leaving a carnival of other cheating services on their platforms.

We will see.

But, to me, the headline of all this is the term I think I just made up — tutorwashing.

It says so, so much that, whoever Einstein is, their first inclination was to turn it into a tutor. To rebrand unapologetic cheating into “the personal tutor every student deserves.” It’s somehow worse.

And it reminds me of the cynical, completely not credible garbage that Chegg pulled for years — insisting with a straight face that they were not defrauding the entire value premise of learning and academic credentials, they were a tutor. They helped students who were stuck. They were better than schools and actual tutors. It was never true. It is never true.

I seriously doubt that anyone at Einstein changed what Einstein does (or did) in hours. I’d bet that Einstein was Einstein, able to do exactly what it was built to do, exactly what they said it did first — cheat. Do the work of learning so students didn’t have to be bothered. The Chegg model. The Course Hero (Learneo) model. The ChatGPT model.

That Einstein thought they could get away with that by calling it tutoring says more than I ever could. From their view, why not? It worked for Chegg for years. It’s still working. Knowing it’s a lie requires skepticism, inquiry, and understanding. Whomever Einstein is, they were not wrong to think they would not get much of any of that. Oh, it’s a tutor? Great!

I’d have some respect for you if you just admitted that a few dollars in your pocket means more to you than whether students actually know anything. Or what a college degree means to anyone. Own your malignance. Hiding behind tutoring is a low con. But it is, quite clearly, the first page in the playbook.

Subscribe now

Universities in India Limit AI Use in PhD Papers — to 10%

According to press from India, The University of Calcutta will limit the amount of AI-created text in PhD work to 10%:

Calcutta University (CU) is set to introduce new regulations for PhD research that will restrict the use of artificial intelligence (AI) in the preparation of research papers and theses. Under the proposed norms, not more than 10 per cent of a PhD thesis will be allowed to be generated using AI.

And:

The university now plans to deploy specialised software to detect the extent of AI usage in academic work. If it is found that more than 10 per cent of a research paper or thesis has been generated using AI, the submission will be rejected.

The article also says that:

Jawaharlal Nehru University and University of Delhi use specialised software to identify AI-generated text, rejecting papers where AI use exceeds 10 per cent.

I don’t have too much to say about this.

On one hand, I’m glad these universities are checking on AI creation, making it clear that they care to know whether the student did not do the work. I’ll never understand schools who don’t check, and by extension, don’t seem to care.

At the same time, the 10% idea is a silly one. If a PhD student did not write 10% of their thesis, do they still get 100% of their degree? Will they get 100% of future wages?

Further, a hard 10% is weird. I can see many circumstances in which 15% AI may be just fine. Others in which 2% may be a real problem. It depends a great deal on what the AI is creating. No? What if the 10% that the AI wrote was the actual research? What if it only wrote the lit review? Or just a summary? I think that matters. And it cannot be captured in a number.

Finally, the article does not say, but it’s probably not too hard to learn what AI detection system these schools are using. Once that’s known, students will pre-check their scores and edit to get a safe 9.3% or whatever. Not only does this not achieve the objective of directing student research away from AI, it wastes everyone’s time.

A student who spends four hours checking, editing, and rechecking their work is engaged in unproductive work — busy work, we used to call it. Pointless. Further, the educators who then scan the revised and reduced work on the same system are wasting their time too. Of course it’s going to say 9.3%. You told everyone it needed to be below 10%.

Most of all, you’ve really only just taught students how to use AI and edit to avoid tripping an artificial standard. I called it silly already. That’s what it is.

Subscribe now

(423) The LMS is Dead

Derek Newton — Tue, 24 Feb 2026 19:41:12 GMT

Subscribe below to join 5,087 (+15) other smart people who get “The Cheat Sheet.” New Issues nearly every Tuesday and Thursday.

As always, The Cheat Sheet is verified to be human-written:

Meet LMS Assassin Einstein

Yes, it’s hyperbole to say the LMS is dead.

But it’s not hyperbole to say there are serious questions about its future use as an academic assessment engine. The day has been coming when agentic AI, automated bots, would render any online, not directly supervised assignment or assessment moot.

That day has probably arrived.

Sent in by a reader, meet Einstein, from a company calling itself companion.ai. The very introduction says Einstein:

logs into Canvas every day, watches lectures, reads essays, writes papers, participates in discussions, and submits your homework — automatically.

Canvas, as you likely know, is a leading learning management system (LMS). And now, if you’re a student and your class uses Canvas, or other LMS, as something like 95% of them do, there’s no point in even showing up — not even online.

ICAI, March 6-8 in Denver

The annual conference of ICAI, the International Center for Academic Integrity, is just days away — March 6 to 8.

If you’re going and want to meet up, please let me know. A reply e-mail to The Cheat Sheet reaches me.

See you there.

Subscribe now

Yale Update

In Issue 420, we shared and covered an NBC story about AI use and AI detection. That NBC offering noted that some students had sued schools when they believed they’d been wrongly accused of cheating.

Although this practice is old — well predating generative AI — NBC uses the legal challenges as a kind of proof that AI detection is unreliable.

One such case comes from Yale, where a student in a part time MBA program for executives sued, claiming he was incorrectly accused of using AI in his work. The student was suspended for one year from the program — not for cheating, but for not being forthcoming with the inquiry, according to press reports. The student sued anonymously, which courts rejected, along with his request for an injunction.

I mention this because the last news on the case is nearly a year old now and I asked Yale for an update — not a comment, necessarily, just an update. I asked twice earlier this month and, as is common, the school did not respond.

Most schools seem to genuinely believe that if they just close their eyes hard enough, cheating won’t exist.

Subscribe now

Class Notes:

A very brief note here on the frequency of The Cheat Sheet.

Some of you may have noticed that it had been a bit since our last issue — two weeks, in fact. I assure you that the gap doesn’t mean much. Two weeks ago, I was out of town for a few days and last Wednesday I had hand surgery. In addition to knocking me off my routine, it’s made typing a slow, tedious challenge. While I recover, I will do my best to keep pace.

As always, should you have news to share, things to say, The Cheat Sheet is open. Please save me from typing.

Subscribe now

(422) Regulators: Regent College London Directed Students to Essay Mills, AI

Derek Newton — Tue, 10 Feb 2026 16:15:28 GMT

Subscribe below to join 5,073 (+12) other smart people who get “The Cheat Sheet.” New Issues every Tuesday and Thursday.

As always, The Cheat Sheet is verified to be human-written:

Regulators Tag Regent College London for Encouraging Cheating

Spotted by a reader, according to coverage in Times Higher Ed, a regulator in England found that staff and others at Regent College London provided support that, by any measure, qualifies as cheating. From the coverage:

Students were given “excessive” support with their assessments including early access to exam questions and were frequently encouraged to submit full drafts of coursework for feedback prior to formal submission, according to a report into the provider published on 5 February.
“Some staff were found to have encouraged students to paraphrase content to reduce similarity scores on plagiarism detection software, and in some cases directed students towards the use of artificial intelligence tools or essay mills,” it adds.

The investigation further found that marks awarded to students appeared to be “excessive and not justified by their performance”.
“The assessment team found that the level of educational challenge was not appropriate to the level of the courses delivered. Too much time was spent helping students pass assessments. This excessive assessment support resulted in a level of rigour and difficulty that was less than the minimum reasonably expected of an undergraduate degree in business and management.”

According to the coverage, Regent College London delivers “franchised courses” to other schools in business disciplines. The regulator is England’s Office for Students (OfS).

Three quick things:

The US has no regulator to do such things. When sub-prime schools do crap like this here — and they do — no one notices and nothing happens. Students are defrauded and the value of a college degree goes down. Every day. Without even the threat of being caught.
In this case, the regulators also called out and/or sanctioned at least two schools that bought courses from Regent — University of Greater Manchester and Buckinghamshire New University. The reason: lack of oversight, essentially allowing a third party to give away degrees under their brand. I love this. Buyers need to be held responsible for the junk they buy.
A quick reminder, from Issue 358, that many so-called credit recovery providers in the United States — companies that offer basic, universal college courses for easily transferable credit to a catalogue of sub-prime colleges — are engaged in some highly questionable practices related to rigor and integrity. In Issue 358, for example, Study.com announced it was not going to proctor its online exams anymore — not ever. Because their students wanted the courses to be easier.

Subscribe now

Professor Uncovers Likely Cheating, Brings in AI-Powered Oral Exams

Panos Ipeirotis is a professor at the Stern School of Business of New York University, and, if you’re an educator or dean, he has a blog post you should read.

It’s good and quite detailed. It’s also surprising.

One surprise is that the good professor does not ban AI use, he encourages it. Nonetheless, Professor Ipeirotis and his teaching colleagues found that the written work of their students was not nearly matching their ability to discuss or explain what they had (supposedly) written. Which means his students were submitting good work but not learning anything.

From his blog:

"pre-case" submissions (short assignments meant to prepare students for class discussion) were looking suspiciously good.

So we started cold calling students randomly during class.
The result was... illuminating. Many students who had submitted thoughtful, well-structured work could not explain basic choices in their own submission after two follow-up questions. Some could not participate at all. This gap was too consistent to blame on nerves or bad luck. If you cannot defend your own work live, then the written artifact is not measuring what you think it is measuring.

Emphasis is his.

Also, no surprise. And this distinction is crucial — allowing AI is fine as a teaching choice, but doing so means you cannot judge submitted work as evidence of learning. At least not alone.

So, the professor decided to use oral exams, powered by AI and an AI voice agent. He records the costs, benefits, and outcomes very clearly, along with advice. It’s really good to review, as an assessment option, in my view.

I’ll also point out that, during the AI verbal assessment, the professor:

asked students to record themselves while taking the exam (webcam + audio). This discourages blatantly outsourcing the conversation, having multiple people in the room, or having an LLM in voice mode whispering answers. It also gives us a backup record in case something goes really badly.

Bold is original.

But imagine that. Even in a class where AI is encouraged, where exams themselves are devised and delivered by AI, the professor was smart enough to know that observation was necessary. Without it, this exam, all exams, are pretty pointless.

Please, if you’re in this conversation about AI in education, or integrity, read his post. There’s meat in it, such as this:

Only 13% (of students) preferred the AI oral format. 57% wanted traditional written exams. 83% found it more stressful.
But here’s the thing: 70% agreed it tested their actual understanding: the highest-rated item. They accepted the assessment but not the delivery.

Of course they hated it. Learning is hard. Cheating is easy. Tells me it’s working.

Anyway, love it. Great work. And it deserves attention.

Subscribe now

EdTech Chronicle Awards

Many of you know that, in addition to The Cheat Sheet and my regular jobs, I publish EdTech Chronicle (ETC).

I mention this because ETC has annual awards, for which the deadline was just extended to early March.

I mention that because one of the categories in the awards is “Best Product or Service Supporting Academic Integrity.” That’s a category because it’s me, it’s mine. And I think it’s important.

So, if you’re reading this and you have a product or service that supports integrity, or you know of one, I encourage you to enter it.

Subscribe now

(421) Cheating on the SAT

Derek Newton — Thu, 05 Feb 2026 18:17:05 GMT

Subscribe below to join 5,061 (-1) other smart people who get “The Cheat Sheet.” New Issues every Tuesday and Thursday.

NYT: SAT Exam Questions Compromised

The New York Times has coverage outlining that questions from the, now digital, SAT appear to have been compromised — showing up in databases designed to help students “practice” the test. Or just cheat.

The headline is more than a little misleading:

Students Are Finding New Ways to Cheat on the SAT

That’s not supported by the story. It is likely true that students are cheating on the SAT with compromised questions and technology hacks. Still, the story presents that questions from the test repository were likely available in the open market, and that it is possible to cheat on the new digital version of the test. That’s it.

If you understand cheating or exam security, neither of these should surprise you.

Once a question is given on a test, it’s burned. In less controlled settings professors have regularly reported their exam questions being available on Chegg and Course Hero within the hour. Consider, from the article:

In some cases, metadata showed that questions were posted online almost instantaneously after the test

Providers of higher stakes assessments such as the SAT have known for a long time that their major concern is theft of exam materials — this very problem. Given what it costs to develop even one question, the issue is a serious one.

The second issue — the possibility to hack the digital SAT exam during the test — should make clear, yet again, that it is not possible to fully secure a digital assessment. Maybe it’s not possible to fully secure any assessment, but whatever the best possible threshold is for an in-person exam, digital methods fall short of it.

From the coverage:

Three years ago, after nearly a century of testing on paper, the College Board rolled out a new digital SAT.

If you’re not up on the story, here it’s important to distinguish between the exam being digital, which means taken on a computer, which it is, and remote, which it is not. Takers of the SAT have to personally visit a test center to take the test. The test however, is now on personal laptops — which is part of the problem.

Continuing:

Students who had long relied on No. 2 pencils to take the exam would instead use their laptops. One advantage, the College Board said, was a reduced chance of cheating, in part because delivering the test online meant the questions would vary for each student.
Now, however, worries are growing that the College Board’s security isn’t fail safe. Fueling the concerns are what appear to be copies of recently administered digital SAT questions that have been posted on the internet — on social media sites as well as websites primarily housed in China.

The College Board said in written statements that SAT cheating is rare, affecting only a fraction of 1 percent of its test scores, and noted that overall test scores have remained steady after the transition to digital tests. “However, some students will always be tempted to cheat on high-stakes assessment, and bad actors are persistent. We stay hypervigilant,” it wrote.

I believe they do.

No exam is 100% resistant to cheating. Still, as mentioned, the likelihood of misconduct went up when the SAT went digital, not down. They had to have known that, despite their statements to the contrary.

Moving along:

In addition to having human proctors present during testing, the SAT’s Bluebook platform requires that other apps on a student’s computer be turned off.
But the organization acknowledged that it was aware of “screenshots that purport to have been taken while testing is in progress” as well as “hardware and software-based efforts to evade our security system.”

Not new. We went through all this with remote test proctoring like five years ago. Motivated people can bypass your digital security.

The NYT also provides some context:

Allegations of SAT cheating follow breaches in digital LSAT and GRE tests used by dozens of prestigious law and graduate programs to screen applicants. In those cases, the tests were taken remotely, not in test centers like the SAT.
Several graduate business schools, including those at the Ohio State University and the University of Minnesota, rescinded admissions offers in 2024 following allegations of GRE cheating by students in Ghana and Nigeria who had taken the tests at home, according to the online publication Poets & Quants, which writes about business schools.

We covered the LSAT thing in Issue 391.

But hold the phone — the GRE allowed students to take it at home? I hope I spelled “flabbergasted” correctly. That’s just — I don’t know. In all seriousness, why bother giving any test if you’re going to allow people to take it at home?

There’s also:

Cornell Law School has notified students that it is investigating a testing breach during a contracts exam in December. The school learned of the problem because a paid test taker — intending to advertise his services — posted screen shots of the exam online shortly after gaining remote access to a student’s laptop.
A spokeswoman for Cornell confirmed that its law school had investigated an “isolated incident of alleged testing misconduct,” adding that the school would not comment on disciplinary cases.

I’m not sure how they know it’s isolated. Feels more like self-soothing wishful thinking.

A bit more:

Rachel Schoenig, a former head of security for the ACT, the second-largest undergraduate admissions test, said the web is filled with ads from companies selling hardware and software, purportedly to help students cheat.
“Some are scams,” she wrote in an email. But, she added, “The reality is that no system is foolproof.”

As mentioned, true. And:

The fact that students are permitted to use their own laptops is one weak link that several experts flagged.
There is evidence that digital test security problems are spreading beyond standardized admissions tests, with companies operating online that offer to take college exams for students, for a fee.

Wait — companies offering to take college exams for a fee? You mean like what I wrote about in 2015?

Anyway, the digital SAT is probably compromised. There’s that.

Subscribe now

Another Article on “Ethical AI” and AI Detection

Terri Smith has written an article published by the OLC — The Online Learning Consortium. OLC says Smith, “has more than twenty years of experience in the education sector. Currently, she is a technology faculty member at a college preparatory school.”

The article from Smith aims to tell educators, yet again, that they cannot trust, and therefore should be reluctant to use, “the AI.” Smith is misinformed and, as a result, misinforms.

The piece is not all bad. I like, for example, this:

Unauthorized AI use raises significant ethical concerns, particularly when it substitutes for independent work.

And this:

Unauthorized AI use … when defined as prohibited, is deceitful and thus unethical. After all, one does not accidentally open an AI chat and inadvertently gain a precisely written answer to an inquiry, and then paste the copied answer into a homework assignment.

Right on both counts.

But most of the rest of the article repeats incorrect or outdated assumptions and shows lack of logic. Smith says, for example:

On occasion, some AI detectors have misidentified student work as AI-created, thus making accountability for plagiarism or cheating problematic.

That is correct — some have. Not all.

But rather than examining, or even considering, how to avoid those inaccurate systems, Smith concludes that because some are bad, none can be trusted. That’s not logical, especially when the solution is so obvious — find and use the good ones. Instead, the piece contends we should stop farming because there are some bad apples.

In another example, Smith writes:

we must ensure that accusations impacting students’ academic records are accurate. But, how can we be sure if the AI isn’t?

We must ensure that accusations are accurate, as much as we can. But accusations themselves don’t mean anything. Findings, adjudications, consequences need to rest in accurate information, lest they be unjust.

Accusations are, for the most part, investigations, an opening of questions revolving around suspicion or curious data. These accusations/inquiries can be stressful for all involved, and significantly burden school, educator, and student alike. So, they need to be based on more than whimsy. Still, on their own, accusations mean nothing. Most are dismissed or lead to as little as a warning. I’m just not sure, in other words, that accusations impact students’ records, as Smith says. I’ll be stronger. I think that’s simply wrong.

Smith is also wrong to use “the AI.” There is no such thing. You cannot — should not — say “some” are bad, then lump them together as “the.” It’s an error made many times, such as:

Accusations of unauthorized AI use are problematic due to the unreliable outcome of the detectors

“The detectors.” And:

… the AI may have constructed a false-positive.

“The AI.” And:

until AI detectors become more reliable

“AI detectors.”

You get it.

There are several other issues. Here’s another:

AI Detectors are the AI tools that analyze the probability of whether or not text is AI-generated from ideas, concepts, conventions, and/or writing structures that appear to have AI patterns. Detectors are not reliable (yet) because the training for the AI models is still developing.

Someone could correct me if I am wrong here — several actual AI experts subscribe to The Cheat Sheet — but I do not believe any AI detection system examines “ideas, concepts and conventions.” I do not believe that either the AI that created the text, or the AI that finds it, understands the ideas and concepts at issue. If you know otherwise, please let me know.

I do know that Smith is wrong when she says “Detectors are not reliable (yet).” Good for using “yet,” and I note again the misleading use of one class of “detectors.”

But in this case, “yet” happened three years ago. There was never any evidence that all AI detection systems did not work. But all of the most recent research shows that AI detection by good, credible systems is both accurate and reliable.

I’ll end with this, from Smith:

Be careful! If a student completely denies AI use, then the evidence to counter the denial may be absent.

It’s a shocking thing to believe that, if a student denies using AI, there may be no evidence. I just don’t know what to say. It’s a premise I struggle to process. In what context does a denial invalidate evidence? That’s downright bizarre.

Since it’s obvious that readers cannot trust the writers of this material to be accurate, or even logical, I really wish that editors at some of these publications would be vigilant. There’s just no excuse for things such as this to find their way into publication and therefore into our conversations. But I wish for too much, obviously.

Subscribe now

Registration Open for ICAI Conference, March, in Denver

A reminder, registration is open for the annual conference of ICAI - the International Center for Academic Integrity.

This year’s festivities will be in early March, in Denver.

If you or your campus may be eligible for one of the ICAI’s Integrity Awards, nominations are due by February 9.

Subscribe now

(420) NBC on AI Detection

Derek Newton — Tue, 03 Feb 2026 19:05:13 GMT

Subscribe below to join 5,062 (+17) other smart people who get “The Cheat Sheet.” New Issues every Tuesday and Thursday.

NBC Writes on AI Cheating, Detection

As these articles go, the recent offering from NBC is not awful. In fact, it’s pretty balanced.

Credit to the writer, Tyler Kingkade, for not chasing heat and sacrificing light. And big credit as well for quoting not one, but two actual experts on academic integrity. Imagine that.

Granted, the headline is awful. And several of the things people were quoted as saying are bonkers. But bonkers is where we are. Let’s start with the headline:

To avoid accusations of AI cheating, college students are turning to AI

Having spent years writing about cheating, integrity, and now, AI, this headline was triggering. I don’t like that word. But it was.

How many fantasy fiction articles have “news” outlets published in which the focus is accusations, instead of actual cheating? So much digital ink has been wasted writing about the accuracy of the radar gun with the few people who insist they personally were not speeding, while suburban streets have turned into the Monaco Grand Prix. This headline prepared me for yet another of these — but I didn’t cheat, the tool is wrong.

Yes, I am aware how old I sound.

This story is that. But again, not one-sided. And most of what is crazy about the article comes from the people in it, not the writer. I respect that.

Let’s go through it.

An early paragraph:

Rapid adoption of AI by young people set off waves of anxiety that students could cheat their way through college, leading many professors to run papers through online AI detectors that inspect whether students used large language models to write their work for them. Some colleges say they’ve caught hundreds of students cheating this way.

True. People were — and continue to be — very worried that students could use AI to cheat their way through college. Because they do (see Issue 417). And schools have said they have caught hundreds, which most people also acknowledge is only a tiny fraction of actual incidents.

Here’s the next section:

However, since their debut a few years ago, AI detectors have repeatedly been criticized as unreliable and more likely to flag non-native English speakers on suspicion of plagiarism. And a growing number of college students also say their work has been falsely flagged as written by AI — several have filed lawsuits against universities over the emotional distress and punishments they say they faced as a result.

You have to read this one closely, I think.

Yes, AI detectors “have repeatedly been criticized” as unreliable. Good work here to not say they are unreliable. Criticism is not fact. The non-native English speaker issue is junk (see Issue 216). For one, that study is from 2023, which is ancient history in terms of the advancement of AI and AI detection technology. Also, as I’ve been screaming about since the study came out, it’s absurd. It has made-up citations for crying out loud. No one seems to care. It’s repeated, and repeated, and repeated.

Also, yes, students say they are being falsely accused. Like criticism, this is not a fact. Saying you are innocent does not mean you are innocent. Saying you’re innocent is, in fact, the expected human response to nearly any accusation.

But if you ran through that paragraph above quickly you may not have caught that AI detection is criticized and that students say they are falsely flagged. Both are true and, in my view, pretty weak foundations on which to build a conversation.

Moving on. There’s this:

Amid accusations of AI cheating, some students are turning to a new group of generative AI tools called “humanizers.” The tools scan essays and suggest ways to alter text so they aren’t read as having been created by AI. Some are free, while others cost around $20 a month.
Some users of the humanizer tools rely on them to avoid detection of cheating, while others say they don’t use AI at all in their work, but want to ensure they aren’t falsely accused of AI-use by AI-detector programs.

The part I’d like to circle here is that companies are selling — literally selling — tools designed to aid students “avoid detection of cheating.”

In the past, I’ve likened these tools to luminol, the chemical that highlights human blood, even after it’s been scrubbed away (see Issue 318). Why anyone who was not trying to hide evidence would have luminol in their house is incomprehensible. I just don’t buy the argument that “I promise I did not kill anyone, but I’m going to check my kitchen floors with luminol — just so no one can falsely accuse me of it.”

Companies that sell humanizers — cough, Grammarly — have compared this practice to insurance — just to be safe (also Issue 318). And it drives me a little crazy that no one is connecting the dots that these companies, the ones selling insurance, literally benefit by convincing people that AI detectors are not reliable. Can’t trust ‘em, better pay us for the insurance.

Again, these companies sell products to help students “avoid detection.” It’s not just that they check for evidence, they help you clean it up. For a fee, of course.

And while we’re here, many of these “avoid detection” companies also have AI-checkers. Cough, Grammarly again. Allergy season. Take a guess — when a company sells the fix, guess what their free check is going to find. That you need a fix, of course. This is the logic they — sneeze Grammarly — really want you to buy:

you can’t trust AI detectors
use ours
trust when ours says you have AI in your document
pay us to clean that up —you’d hate to get caugh… err, be falsely accused. Wink.

That is not serious. They are not serious.

Anyway:

In response, and as chatbots continue to advance, companies such as Turnitin and GPTZero have upgraded their AI detection software, aiming to catch writing that’s gone through a humanizer.

They’re not the only ones, but yes. Not only have these companies upgraded, they continue to do so. Now, if you’re using a humanizer, thinking you’re getting away with using AI, you’re going to be caught — using the humanizer. It’s AI. By using the humanizer, you’re using AI, actually putting AI text in your work. Even if it was not there before.

Have to share this too:

“Students now are trying to prove that they’re human, even though they might have never touched AI ever,” said Erin Ramirez, an associate professor of education at California State University, Monterey Bay.

Just one word: reCAPTCHA.

I’ve never touched a robot, and I have to check this stupid box like 20 times a day:

Welcome to our world, Professor Ramirez. Proving you are human is now something we have to do all the time.

Continuing:

“If we write properly, we get accused of being AI — it’s absolutely ridiculous,” said Aldan Creo, a graduate student from Spain who studies AI detection at University of California San Diego. “Long term, I think it’s going to be a big problem.”
A teaching assistant in a data science course accused Creo of using AI to write a report in November. Creo explained to the TA that he has a habit of explaining step by step how he reasons through a problem, which ChatGPT is known to do, according to a copy of messages he exchanged with the TA.

Further:

to avoid another battle, Creo said he sometimes “dumbs down” his work by leaving words misspelled or using Spanish sentence structures that aren’t proper in English.

Something about this story feels off to me. It may just be that if you think ChatGPT is proper writing, that’s a problem. Its output is mediocre, at best. And if you’re having to write below its level to not be flagged, again, that’s a problem.

Also, I’m not so sure that good AI detectors flag writing “properly” as AI. Maybe someone who knows can clarify, but I think that’s a myth about how the systems work.

I apologize in advance for sharing such a large section next:

At their worst, the stress from the accusations has driven some students to drop out of school.
Brittany Carr received failing grades on three assignments she completed as a long-distance student at Liberty University, a private evangelical school in Virginia that has one of the largest online enrollments in the U.S., because they were flagged by an AI detector. She showed her revision history, including how she’d written one first by hand in a notebook, according to screenshots of emails and messages she exchanged with her professors.
“How could Al make any of that up?” Carr wrote in a Dec. 5 email. “I spoke about my cancer diagnosis and being depressed and my journey and you believe that is Al?”
Her evidence wasn’t enough — the social work school still told her she needed to take a “writing with integrity” class and sign a statement apologizing for using AI, emails show.
“It’s a very weird feeling, because the school is using AI to tell us that we’re using AI,” she said.
It stressed her out. Carr worried another cheating accusation could cause the Department of Veterans Affairs to take away her financial aid. In order to avoid more false accusations, she said, she ran all of her material through Grammarly’s AI detector and changed any section that it highlighted until it concluded a human wrote the whole thing.
“But it does feel like my writing isn’t giving insight into anything — I’m writing just so that I don’t flag those AI detectors,” she said.
After the semester ended, Carr decided to leave Liberty. She’s unsure where she’ll transfer.

First, Liberty. A major suspension of disbelief is needed here if we are to draw any conclusions from this reporting. Liberty is, let’s say, a deeply atypical educational experience.

It sounds as though this student had evidence of her doing the work. But if Liberty’s AI detector flagged three separate examples of her work, something is wrong.

She may be using AI to “polish” her work, not knowing it’s AI. This does say she was using Grammarly, which can rightly trip AI detection. Although we’re kind of told she was using Grammarly after the accusations. It does not say whether she was using it before.

It’s also possible that Liberty’s detection system is deeply flawed. Since it’s not identified, I can’t speculate as to its quality. If schools are cheap on their systems, they will get cheap results. Although I am glad Liberty is at least checking for AI.

Also, AI can absolutely make up your personal experiences. Especially if you ask it to, or tell it what those experiences are.

And finally, I note that this student does not say she left Liberty as a result of the AI accusation issues, as the story said.

Moving ahead:

At the root of many conflicts about students using chatbots to cheat is a disagreement about what counts as too much AI use on homework.

That’s not great either. Lines about quantity or quality of disallowed AI use in academic settings is not a disagreement. Educators set standards for their classes based on a variety of factors, which is how it should be. The idea that there should be one clear standard that fits all subjects, all students, all teaching formats is fantasy.

This too is important:

Independent analyses of AI detectors show mixed accuracy. One pre-print study last year found GPTZero is good at finding AI-generated writing, but “its reliability in distinguishing human-authored texts is limited.” However, other research pegged the company’s detector at near-perfect accuracy. Meanwhile, separate studies from 2023 and 2024 have found that Turnitin had a low false positive rate, but failed to identify more than a quarter of AI-generated or AI-rephrased texts.
Both companies emphasized that research showing flaws in their detectors is outdated due to the rapid evolution of large language models and updates to their own detection software.

Mixed accuracy is fair. Some are terrible. Some are shockingly good.

I’ve not been shy about pointing out that, based on what I’ve seen, GPTZero is not good. The tool matters. And yes — most of the studies people use to try to discredit AI detection are old and, although the article does not say this, they also average the results into a single number, dragging down the performance of the good systems.

Turnitin tells schools never to use its tools as the sole basis for deciding whether a student cheated, said Annie Chechitelli, the company’s chief product officer. It should instead prompt a conversation with a student about how and why AI was used, she said.

Yes, again. Every quality AI detection system advises educators not to use the score alone as proof. It’s a piece of evidence, not the only evidence.

Professor Ramirez, from above, also:

said anyone who relies on a detector has never put their own work through it.
“It’s almost like the better the writer you are, the more AI thinks you’re AI,” she said. “I put my own papers into AI detectors just to check because I don’t like to hold students accountable without knowing how the tool works. And it flags me at like 98% every time, and I didn’t use AI in any capacity.”

I call BS on this.

Again, if you think your writing is being flagged as AI because it’s too good, you are somewhere else in this conversation. Also, she put her work into “AI detectors,” more than one. Which ones? Grammarly? That really matters. If she used a good one, there is no way — zero chance — that 98% of her writing is flagged as AI “every time.” Simply not to be believed.

Another longer section, sorry:

Turnitin, which has been around for a quarter-century offering tools to help educators catch plagiarism, is trying to keep up with humanizers. The company views humanizers as a “growing threat to academic integrity,” and issued a software update last August to detect text modified by the tools.
The company has a list of 150 tools that charge as much as $50 for a subscription to adjust text so that it’s not flagged by an AI detector. Chechitelli referred to them as companies whose “sole goal is to really help students cheat.”
Demand has surged for the tools. Joseph Thibault, founder of Cursive, an academic integrity software company, tracked 43 humanizers that had a combined 33.9 million website visits in October.

Nothing to add beyond that Chechitelli is right. The goal of these companies is “to really help students cheat.”

Speaking of:

Superhuman, the company that makes Grammarly, developed a tool it calls Authorship that’s included with basic accounts. Students can turn it on to surveil themselves on Google Docs or Microsoft Word as they write and playback later. It will show which sections were typed, pasted from another source or generated with AI.
“We’re going to keep track of when you are going to Wikipedia,” said Jenny Maxwell, Superhuman’s head of education. “We’re going to keep track of when Grammarly is making suggestions and you’re taking them, we’re going to keep track of how much time you’ve spent in this paper or how many sessions.”
As many as 5 million Authorship reports were created in the past year alone, she said, though most of the time they aren’t submitted.

Most are not submitted. That’s interesting.

If students are not using AI — or are using it appropriately — and they’re beset upon to have to prove they’re human, enduring the stress of showing the work of learning, why are most of these reports not submitted?

I need to wrap this up. But one final thing. The NBC story goes from Grammarly to the TikTok famous Georgia student who claimed she was falsely accused of using AI when all she did was use Grammarly (start at Issue 282, you may have to work backwards).

That student, who graduated by the way, said it:

increasingly became hard to avoid writing on software that didn’t have it embedded.
“Google has AI embedded into it, Microsoft has AI embedded into it — like literally everything has AI in it,” she said. “So, in a roundabout way, there’s no way to write a paper without using AI, unless you go to the library and you check books out and use encyclopedias.”

Read that again. Please.

There is no way to write a paper without using AI, unless you go to the library, she says. In addition to being false — just flatly wrong — this stuns me. The former student and Grammarly social media advocate who insists she did not use AI to write her papers, says there is no way to write a paper without using AI. Got it.

What are we doing here? By any measure, this is insane.

I’ll end where the NBC story does, with a quote from integrity expert Tricia Bertram Gallant, who rightly says that not enough pressure or attention is being put on the companies that created these AI products — the products millions of students are using and misusing in education. She says:

“We keep turning on what the academic institutions need to do to fix problems that they didn’t create,” she said.

No joke.

Subscribe now

Another Update on The University of Chicago

In Issue 417 we covered a University of Chicago graduate who told the world that he used AI to cheat his way to his degree.

In Issue 419, I wrote a follow up that despite a few requests, the school had not responded to whether it used any AI-detection technology or whether it was possible to cheat as much as this student claimed, and graduate.

After writing the follow up in 419, the school did respond. On January 27, Gerald McSwiggan, Director, Public Affairs wrote to me. This is his entire missive:

Derek,
Thanks for the email.
Please see this message that President Alivisatos sent to students at the beginning of the academic year:
https://president.uchicago.edu/from-the-president/messages/ml-ai-and-your-education
Best regards,
Gerald

Left in the link so you can see that I was directed to a written, “Dear Students” statement advising students to be ethical, essentially. It’s a highly recommended read for its sheer obtuseness. Here’s one example:

Today you can ask questions in a genuinely new way; you can “just” ask a machine, an AI of one kind or another, which seeks to leverage past knowledge in service of the new.

Golly.

Here’s another:

Are these shortcuts that forestall your learning? Are they powerful tools that let your intellect soar higher? Yes and yes.

“Your intellect soar higher.” Seriously. Also:

The faculty demand that you approach the use of machine learning and AI-based tools in ways that are ethical and urge you as well to be both skeptical and ambitious.

Emphasis is original.

And the faculty demand that AI use be ethical. That ought to do it. Good job.

Before you think I missed the point, here:

Here is the take-home message. You came here to learn how to think. Today that means learning to think with machines and learning to think without them. If you honor that approach, yours will be the generation that indeed soars well beyond the levels of knowledge previously imagined.

Again, emphasis original. And deeply silly. And soars again. Seriously?

But the message — the take-home message — is, I guess, that you can soar if you honor that you came to the University of Chicago to learn how to think. I mean — cool. But, what?

It does not even say don’t cheat. It says to be ethical. Sure. I am confident that turned the trick. It reads like a middle school student found the font keys.

Getting out of that bubble of whatever that was, I’m returning to the response from the University. I asked a simple yes or no question and for a comment, should they want to add one. That’s what I got. A bunch of soaring and crazy emphasis choices.

I asked if I needed a license to fish in the river. I got three copies of Field and Stream from 1987.

I repeated my questions. I’ve received no reply.

The lack of regard for this issue from the University of Chicago is stunning. It’s the only word I have.

Subscribe now

(419) Growing Misconduct Cases at LSU

Derek Newton — Tue, 27 Jan 2026 19:18:32 GMT

Subscribe below to join 5,045 (+6) other smart people who get “The Cheat Sheet.” New Issues every Tuesday and Thursday.

AI-Related Cheating Cases Growing at LSU

A local TV news station has a story about growing cases of AI-related misconduct at Louisiana State University.

It’s curious that the story does not seem to give us any actual data about the number of cases or growth rates. It focuses instead on the backlog of cases and students who, quite predictably, claim that the allegations and process are unfair. Nonetheless, it seems clear that cases of misconduct allegations are increasing.

Here is the story’s first paragraph:

LSU students are facing a surge of artificial intelligence cheating allegations that has created a significant backlog at the university’s Student Advocacy and Accountability Board, according to documents obtained by WAFB and student accounts.

A surge, we are told.

The story quotes a student identified as Sarah, who says:

“I went to go check my grades, and I saw that I had a zero. So I went, and I checked to see her notes, and my teacher completely scratched all of it out and said that it was 93% AI written,” Sarah said.

Seeing 93% here tells me the professor was probably using an AI detection system and, given that they shared the score, it’s possible that they are using this system and nothing else. That’s bad practice. Granted, 93% could be strong evidence. Still, bad practice.

Also, while we’re here, a zero on an assignment is not a penalty — it’s an earned mark for not having done the work. Yes, my opinion. Happy to defend it.

The story also has this:

With scholarship money at stake, Sarah’s case was not resolved until January. She said she was given the option to appeal or admit to using AI.
“I said that I used AI because one of my scholarships needed my grades, and if I appealed that I didn’t use AI, it would just prolong the process, and I really needed to submit them my grades,” Sarah said.
Sarah said her father wanted to fight the allegation with a lawyer, but the family decided against it due to concerns about losing scholarship funding.
“It wasn’t fair that my school tuition was on the line,” said Sarah.

I am sympathetic to case backlogs — on both sides. It’s not fair that students must wait on case dispositions, especially when lack of a finding impacts other important things. And I don’t like that an inefficient process can pressure students to essentially surrender.

It’s also quite understandable that schools have backlogs. They are not, in most cases, equipped to deal with these situations, and especially not at scale. Hundreds, thousands of cases would cripple any system. And most schools are really, really resistant to investing the time and money it takes to handle these cases adequately and efficiently. Plus, as evidenced in this example, some of these cases can be litigious; they require evidence and adequate process and preparation.

Two other quick things on the blurb above.

One, it absolutely is fair that tuition is on the line in cases of alleged cheating. Cheating is academic fraud. Continued standing in your academic community very well ought to be on the line. I do not understand the thought that connects a possible incident of academic fraud with the assumption that accused students gets to keep their scholarship. Should they be held responsible, of course.

Two. “I really needed to submit them my grades.” I just cannot.

Back to the article, the closest we get to data is:

According to public records from LSU, dozens of students are dealing with similar situations.

I know it’s TV news but if we’re talking about dozens, and that’s a surge, I have to wonder where we started.

The article shows an e-mail from a student to a Dean about cheating allegations. It includes:

Everyone either has a zero or no grade in the gradebook, people are crying, and we are being told that it is our faults.

I have so many things about this.

This is also from that student e-mail:

Personally, I did not use Grammarly, nor any other form of AI and I too am being wrongly accused of AI usage.

Interesting that this student seems to know that using Grammarly is AI usage and may be flagged as cheating.

There’s also this, because of course:

Professor Andrew Schwarz, who works in the Entrepreneurship and Information Systems Department at LSU’s College of Business, said AI detection systems have limitations.
“An AI system cannot determine whether or not something that is generated is AI or not,” Schwarz said.

Professor Schwarz is wrong, as we’ve documented over and over again. But recalling one of my favorite phrases from childhood, you can’t tell a smart boy much. Schwarz says:

students are experiencing anxiety about AI use policies.

I quit.

For their part, LSU released a statement to the station:

At LSU, high standards of academic integrity are essential to our mission. As AI-related academic misconduct referrals continue to rise, each case is reviewed under the Student Code of Conduct to ensure that grades reflect a student’s own work and that any concerns are addressed through a fair and established process.
If a Student has a question regarding the Instructor’s expectations for assignments, projects, tests, or other items submitted for a grade, it is the Student’s responsibility to seek clarification from the Instructor.

Good for LSU.

They acknowledge that referrals are increasing. They underscore the need to establish that grades reflect student work and say, quite rightly, that it’s a student’s job to seek clarity if they are confused.

Consider please, for a moment, the “real world” of a career. No employer is going to pay you for work you do not do. Not for long. And if you don’t understand the rules, you need to ask. No one in a professional setting is going to accept that the rules are confusing as an excuse.

Anyway, good for LSU.

Subscribe now

The Independent Fumbles

The Independent, a paper in the UK, has a really bad — albeit entirely predictable — story out recently about AI and misconduct.

Short version: poor students, being falsely accused of cheating by AI detectors. Yes, we’ve been fed this a thousand times. No, it’s still not true. At least not a trend or state of affairs.

I wish I could leave it there.

The story says:

Nearly half of (43 percent) U.S. teachers with classes from sixth to twelfth grade said they used AI detection tools in the 2024/2025 academic year, according to a recent poll by the Center for Democracy and Technology.
However, while accusations of cheating with AI can lead to docked grades, probation, or even expulsion, they can also have a serious impact on the students themselves, experts say.

I’m not sure 43% is nearly half. But whatever. I am sure that “accusations” of cheating don’t lead to all those things. Not on their own. To have grades docked, be suspended or expelled, usually requires a finding of responsibility — a determination that, in fact, the cheating took place. The Independent skips that part.

But here we go anyway:

“One of the most common feelings that students describe to me is anxiety and stress from even going through the process, even if they’re saying I'm innocent,” Lucie Vágnerová, a New York-based education consultant with over 10 years of experience, tells The Independent.

Yes, being accused of misconduct is stressful. It causes anxiety, I am sure. Tell me how to manage academic integrity and enforce policies against cheating without creating stress. I’m all ears. But we go on:

“A lot of them tell me they are not sleeping well – a lot of them have to seek out counseling, and the misconduct process at U.S. colleges and universities often takes at least several weeks, if not months, sometimes, so this is really like a long-range situation, really deeply affecting their mental health.”

I am not saying any of that is good. But tell me how to avoid it. Moreover, this is not a reason to impugn integrity policies, or enforcement thereof.

The Independent retells the joke of a story about the University of North Georgia student who says she used Grammarly but was found responsible for misconduct, lost her grades and scholarship. If you’re interested, start at Issue 282 and link backwards.

Experiences such as these ultimately erode trust in the educational process and ultimately the core relationship between students and their teachers, says Vágnerová.

Cheating does that.

And still more:

“To me, that's a huge problem,” [Vágnerová] tells The Independent. “Institutions are investing in all this surveillance, and they are not investing in instructors’ ability to build deep relationships with students and build that trust and that vulnerability.”

I agree about paying teachers more. But otherwise, think we can all stop listening to her now.

But unfortunately, we’re not done. The article goes on:

Research has also found that such detection systems are, in the best-case scenarios, limited, and in the worst-case, totally unreliable.
“Detection tools for AI-generated text do fail, they are neither accurate nor reliable,” a study by members of the European Network for Academic Integrity found, noting that all the tools they evaluated all scored below 80 percent.

I left the link so that you may check my math here, if you like.

The linked open-access study is the 2023 work from Debora Weber-Wulff and others that we covered thoroughly in Issue 250. It has more holes than Swiss cheese in The Vatican cafeteria.

Recall that, in this very paper, the 14 tested detection systems were 96% accurate at identifying human-written text:

The overall accuracy for case 01-Hum (human-written) was 96%.

This too, is a finding from that paper, pulled from Issue 250:

Four of the 14 systems (CheckforAi, Winston AI, GPT-2 Output, and Turnitin) were 94% accurate with the AI work, missing just one of 18 test samples each. Another two detectors (Compilatio and CrossPlag) were 89% accurate with AI text.
Three of the four that were 94% accurate with AI were also 100% accurate with human-written work. So was one of the systems that was 89% accurate with the AI.
In other words, three of the 14 tested detectors were more than 96% accurate overall — getting the human work perfect and missing just one of 18 AI works. Another was 92% accurate overall - getting all the human work right but missing two of 18 AI submissions.

But they’re — how did The Independent describe them? — limited and unreliable. Mates, that’s just made up. Here’s an idea: read the study. Here’s another idea: interview someone who can tell you about it.

Also — and this really, really annoys me — the study’s authors are members of the European Network for Academic Integrity, I guess. But that organization, as far as I know, played no role in that research and has not endorsed it. That’s red letter misinformation.

I’m a member of the Education Writers Association. They have no idea what I am writing here, nor should they. To cite “The Cheat Sheet” as being written by a member of the Education Writers Association may be true, but it’s deliberately misinforming readers — attempting to convey credibility where none exists.

Not to mention, this study is from 2023 — which is a thousand years ago in AI time.

The article also cites not one of the more recent scholarly studies showing that AI detection is highly, highly accurate. You can find a few of those in Issue 390, even though it seems that The Independent cannot.

The story ends with more nonsense from Vágnerová, making it clear that this is the only person The Independent interviewed:

“I think a lot of people would imagine that the solution is more accurate surveillance,” she tells The Independent. “ To me, that’s sort of the opposite.”
“I think there is a role for AI detection in the education space, but it’s a much, much smaller role than it has now. Ultimately, I think institutions and governments need to invest in compensating educators so that they have the space to create assessments that evaluate student growth meaningfully.”
The ENAI study authors agree.

No. The ENAI study authors do not agree. One, you did not ask them. That paper says not one word about teacher pay or the space to create assessments. Two, this is not an ENAI study.

I’m horrified and embarrassed that something this bad, this biased and misleading, made it to print.

Subscribe now

A LinkedIn Post of Note: 5 AI Cheating Stats That Made Me Uncomfortable

Juliette Denny, the “Founder of Iridescent Technology and Growth Engineering | Host of Human After All Podcast,” has a LinkedIn post that I thought was worth sharing.

I have not verified her sources here or checked to see that she relayed them faithfully. I wish I had that time. Nonetheless, I see no reason on their face to discount them, and Denny does link to her sources, should you have the time to dig around.

She also gets a few things clearly wrong but, it’s really hard to write about this topic. And she’s not a professional journalist. So some slack is cut.

Nonetheless, I’ll share a few of her five stats, which she says made her uncomfortable:

because they named something I’d been trying not to look at directly.

It’s also good context that:

I run an AI learning company. I believe in this technology. I think it can genuinely help people learn. But I also think we need to be honest about what’s actually happening in classrooms right now

With that, here’s one of the five:

Nearly one in five students, 18%, are submitting completely unedited AI-generated work. Not paraphrased. Not edited. Not even proofread.
Copy. Paste. Submit.
That stat comes from a 2026 study by AnsonAlex.com, cross-referenced with data from Stanford and Pew Research. And honestly? I keep thinking about what that means. Because that’s not a student who struggled with an assignment and asked AI for help. That’s a student who has completely disconnected from the idea that their own thinking matters.

I’ve nothing to add.

Another:

58% of students use AI as a tutor for things they’re too afraid to ask about in class. That’s from the same AnsonAlex study. They called it “Socratic Tutoring.”
Read that again.
More than half of students are so scared of looking stupid in front of their teacher, or their classmates, that they’d rather ask a machine. And I don’t know what to do with that information. Because on one hand, yes, AI is helping them get unstuck. It’s giving them a safe space to ask questions without judgment.

I submit that, for at least some of these students, saying AI is a tutor because they are afraid to speak up in class is the safe answer. The cynic in me cannot dismiss that the real reason is that AI will spit out the answer when a teacher may ask them to do the work of learning. Fear is an acceptable, even empathy inducing, explanation.

Still, 58% is a high, high number. For any reason.

Another:

Here’s the kicker: 94% of AI-generated assignments still go completely undetected. Even though discipline for AI-related misconduct increased by 33% between 2022 and 2026, almost all of it slips through. (That’s from AllAboutAI.com, analysing Turnitin and BrowserCat data.) We’ve built this entire enforcement apparatus, detection tools, policies, academic integrity boards, and it’s catching almost nothing.

Do I know that 94% stat is accurate or can be trusted, no. Do I buy it anyway? Yes. Why? Because no one wants to admit this problem, because fixing it is hard.

Subscribe now

University of Chicago Follow Up

In Issue 417, we covered a University of Chicago graduate who was trying to give his economics degree away because, he says, he used AI to cheat on nearly every assignment — “on literally every single essay,” he said.

In 417, I noted that I would ask the University of Chicago if it used AI detection technology, or if it had a comment about whether someone could graduate having used AI in the way this student says he did.

I did.

I asked Alexia Elejalde-Ruiz, Director of Media Relations, for comment on January 13. I followed up on January 20. Then, yesterday, I forwarded my questions to Katherine Ritchey, Assistant Vice President for Communications Strategy and Chief of Staff.

Silence.

The school’s student handbook says:

It is contrary to justice, academic integrity, and to the spirit of intellectual inquiry to submit another’s statements or ideas as one's own work. To do so is plagiarism or cheating, offenses punishable under the University's disciplinary system. Because these offenses undercut the distinctive moral and intellectual character of the University, we take them very seriously.

Very seriously.

Maybe you’ll excuse me for believing this is not true.

Subscribe now