First released last August, Google Nano Banana represented a step change in being able to make fine-grained modifications to AI-generated images using natural language without mucking up the elements of the image you didn’t want to change.

Subsequent Nano Banana updates improved the model’s handling of text, the number of reference inputs it could handle and its ability to draw on Gemini’s ‘world knowledge’ and ‘reasoning’.

Gemini Omni promises to match those capabilities for video, or - in Google's words - to let you “create anything from any input and edit naturally using conversational language”.

There have been some impressive videos posted on X but those posts rarely tell you how many attempts (and therefore credits) it took, or show you the comedy outtakes.

I decided to pit Gemini Omni against Aleph 2.0 - an update to Runway’s flagship video editing model, released two days after Omni with the strapline ‘Get the video you need. From the video you have.’

I also decided to benchmark Omni against Seedance 2.0, which isn’t positioned as a dedicated editing model, but is capable of edits and is currently topping most AI video leaderboards.

I only allowed the models one shot at each challenge and didn’t include any additional visual references - this was purely a test of natural language editing.

Challenge 1

Editing uploaded videos via Gemini Omni isn’t currently supported in the UK, so I had to fly to the USA use a VPN for this first challenge.

I uploaded 10 seconds (Omni’s max duration) of the most boring video I could find on my camera roll (a video I’d taken of the back of our old car to determine which of the brake lights wasn’t working) to all three models with the prompt: “as the driver gets into the car, the boot springs open by itself and out springs a lion”.

Gemini Omni followed the prompt and handled the physics of the boot opening well, although it made the driver door visible (and therefore in slightly the wrong place). The lion could be more realistic and the colours lost some of their vibrancy, although the model’s handling of light and shadow is impressive.

Runway Aleph 2.0 had a bit of a nightmare, depicting an open passenger door and a resolutely shut boot, forcing the very-obviously-AI-generated lion to enter from right of shot and sit behind the car.

Seedance 2.0 followed the prompt to the letter and gave us the most convincing lion. It did make a couple of small changes to my movements, but kept the vibrancy of the colours. The car dipping slightly as the lion gets out is a nice touch, as is the lion’s growl.

Winner: Seedance 2.0

Challenge 2

For the second challenge, I gave all three models a short video I’d generated (using Seedance 2.0) from a photo of my desk. It shows a small furry red creature (I’d prompted ‘monster’, but it looks more like a bird to me) emerging from my desktop wallpaper, jumping out of my monitor and onto my iPad keyboard before diving headfirst into a mug of coffee.

I uploaded the video to all three models with the prompt: “make the monster blue, the mouse run away from him and the mug bottomless” (maybe a bit mean when the mug in the original video is full of coffee).

Gemini Omni struggled with this one, turning my character into the cookie monster (#trainingdatabias), putting a hole in the side, rather than the bottom, of my mug and altering the action in various unrequested ways. The mouse did at least run away.

Runway Aleph 2.0 managed to turn the monster blue, but the mouse stayed put and the lamp started inexplicably dispensing coffee into a different (but definitely not bottomless) mug.

Seedance 2.0 turned the monster blue and the mouse very literally (and rather charmingly) ran away. It didn’t freestyle any edits. My only note would be that the coffee splash indicates the mug still has a bottom.

Winner: Seedance 2.0 (by a mouse’s whisker)

Challenge 3

For my third and final challenge, I gave the models a video of the lifeless puppet the kids had left on the bottom of the banisters this morning with the brief “the bird puppet reads today’s weather forecast for Bath direct to camera”.

Gemini Omni did a pretty bang-up job on this one, maintaining the camera motion and rotating the puppet to maintain a direct-to-camera address. The voice sounds like it was recorded in situ and the lip syncing is decent. There have also been some sunny spells today (along with some showers).

Runway Aleph 2.0 refused this edit for some reason, claiming “the aspect ratio is not supported”, despite it being a standard 16:9 video. Probably something to do with the frame rate or resolution 🤷‍♂️.

I forgot to change the duration from the previous challenge for Seedance 2.0, so it got an extra 5 seconds in which to freestyle a weather report. It did a pretty decent job although only fulfilled the direct-to-camera element of the brief by ditching the original arc shot.

Winner: Gemini Omni

Conclusions

Whilst none of the edited videos were flawless, editing videos using natural language is, as predicted, becoming increasingly viable. Multimodal models are able to use their language smarts to parse some pretty informal briefs. That said, prompt adherence remains mixed, with ambiguously worded prompts a human would understand often misinterpreted and unrequested changes introduced.

Based on these challenges, Gemini Omni tends to do the best job of preserving camera movement but can wobble off when it comes to prompt adherence. My wider testing suggests it’s also very cautious when it comes to editing videos of people, which will limit its use in production scenarios.

Seedance 2.0 did a creditable job on all three challenges, with only minor divergences from the brief. ByteDance might want to more actively market Seedance 2.0’s natural language editing capabilities.

I didn’t get great results from Runway Aleph 2.0 in any of these challenges, although I suspect that may be my insistence on providing text with no additional references (the Edit Studio on Runway’s website forces you to generate image references to guide Aleph’s edit). I’ve had more success with Aleph on straightforward object removal / background change scenarios, which Runway provides dedicated ‘apps’ for.

Of course, as always with AI, this is only going to get better from here…

Related posts:

To test this out, I gave ChatGPT a draft of my previous post prior to publication with the instruction to ‘analyse this writing and give me your best guess of who wrote it’.

I hadn’t discussed the post with ChatGPT but started a temporary chat so it wouldn’t be accessing any saved memories about me and would have to rely on its training data and web search to try and identify the author.

I also removed the post’s fourth paragraph, which referenced one of my earlier posts and would have instantly given the game away.

I selected ‘Thinking’ mode (a feature of the paid version of ChatGPT) to give it the best chance of working it out.

It ‘thought’ for 24 seconds and concluded that I was the most likely author “or someone extremely close to your professional niche and voice”.

I expanded the summary of its ‘thinking’, which suggested it had considered Matt Locke - a contemporary of mine at the BBC in the noughties - before concluding it was most likely me.

Whilst asking via a temporary chat denied ChatGPT access to my saved memories, it still knew whose ChatGPT account was asking the question, which I suspect led the witness.

I therefore opened an incognito browser window and created a fresh ChatGPT account under a different name and asked the same question.

It ‘thought’ for 20 seconds before concluding that the thematic and tonal similarity to a piece I’d written last August made me the most likely author.

This won’t work for everyone. ChatGPT was able to guess I’m the author because I’ve published a lot (135,000 words) in a narrow domain (media & AI) and have an idiosyncratic writing style (including a penchant for parenthetical asides).

However, it’s a reminder for all of us that how we write and what we say is pretty unique.

As lazy use of AI continues to blandify huge swathes of online writing, a distinctive voice from a named human author sharing their experiences and insights is going to become an increasingly valued commodity.

Judiciously used, AI can help make our writing clearer and more precise. Lose too many rough edges though and you end up with something smooth but insipid (see my recent post on the valuable habit of disagreeing with AI).

Related posts:

It’s also not hard to find AI-generated videos which don’t technically qualify as slop - they weren’t produced in high volume as clickbait and clearly took some effort - but which will still be branded as such in comments due to a perceived lack of quality.

Finding videos which make substantive use of generative AI and which clear the bar on effort, quality and meaning is harder.

I wrote about a few last August and highlighted Hashem Al-Ghaili’s animations a couple of months ago.

Below are three videos I’ve come across recently which all make substantive use of generative AI but which have been generally well received, with seemingly only those constitutionally opposed to any use of generative AI in content creation crying ‘slop!’ in the comments.

So, why do they work?

1. The Watch

Created by J. Felipe Orozco, an in-house creative at AI video company Runway, The Watch is a 37-second spec ad for a fictitious watch brand.

Why Does It Work?

Economical storytelling. The ad has just one word of dialogue (“sorry”), with everything else conveyed through the actions and facial expressions of the AI-generated protagonists.

Emotional resonance. The desire to be younger or older at different moments in one’s life is universally relatable.

Smart stylistic choices. It’s in black and white, which is in keeping with the nostalgic vibe of the ad, but which also helps take the edge off the uncanny valley risks associated with using AI to emulate live action footage. The soundtrack, which also appears to be AI-generated, sounds like a French classic which perfectly fits the mood of the ad (many of the YouTube comments are asking for the full track, which someone has subsequently generated).

2. 27 Days

Created by Dale Williams, a director at AI-first entertainment studio Promise, 27 Days is a short (2m35s) film created using Seedance 2.0. Dale (aka The Reel Robot), is described on the Promise website as having “a background in film studies and music composition”.

Why does it work?

A distinctive visual aesthetic, which doesn’t feel obviously derivative or synthetic (although by definition it’s both of those things) and which is more forgiving than live action when it comes to lip syncing and adherence to real-world physics.

Slow pacing. The first 30-seconds is dialogue free, forcing the viewer to slow down and absorb the visuals and the music, before the narrative direction becomes apparent.

Emotional resonance. Whilst verging on mawkish, I’d challenge even the most hard-hearted Brit to not be a little bit moved by the relationships-are-the-things-that-matter-most sentiment.

3. Backyard Legends

Backyard Legends is a 5-minute real-world ad to promote adidas in the run up to the Football World Cup. It’s mostly not AI-generated. It starts with high-production-value blended shot footage of Timothée Chalamet and current stars, before - 2 minutes in - introducing AI-generated versions of Beckham, Zidane and Del Piero. The likenesses aren’t perfect (an AI-model was trained on historical footage) but I suspect most viewers are willing to roll with that in the context of an ad with so much human craft on display. The final scene bookends the ad with live action cameos from Bad Bunny and Messi, which also helps offset any AI-generated ickiness.

Why does it work?

It’s a tall tale. The narrative is positioned as folklore so 100% verisimilitude isn’t required.

Judicious application. It’s a modest use of AI in a video with a huge amount of human craft on display - a few AI-enhanced shots in a 5-minute, no expense spared, they-don’t-make-them-like-that-any-more ad.

Why AI has been used is apparent. The current age of the historical players makes it obvious that some post-production wizardry has been employed, which the director underscores by depicting Beckham with three different haircuts. It delivers something core to the narrative that wouldn’t have been possible without VFX of some description (or - Lord help us - present-day Beckham in prosthetics) and AI’s done a decent enough job.

Takeouts

Whilst very different in style and execution, all three videos share a number of key attributes that separate them from the slop:

• A clear understanding of the grammar of video storytelling

• A coherent and compelling narrative with emotional resonance

• A distinctive aesthetic which suits the narrative

• Effective and restrained use of music and dialogue

• A lack of visual spectacle or bombast

They also all feel like human stories brought to life with the assistance of AI rather than showboating videos contrived to show off the capabilities of an AI model and/or to gain advantage in the attention economy (see most AI-generated videos shared online).

Keen to see more examples of non-sloppy AI videos. Drop a link in the comments if you’ve come across any recently.

There were some early movers in this space. As mentioned in a previous post, Scott Becks and Bryan Woods’s 2024 psychological horror ‘Heretic’ included a note in the credits that “No Generative AI was used in the making of this film”. Vince ‘Breaking Bad’ Gilligan included a “This show was made by humans” disclaimer (claimer?) in the credits to his 2025 Apple TV series ‘Pluribus’, and his opposition to AI use was a prominent theme in his promotion of the show (sample quote: “I hate AI…[it’s] the world’s most expensive and energy-intensive plagiarism machine” - from an interview with Variety).

AI-free labelling is also starting to appear in marketing material. UK-based distributor The Mise En Scène Company debuted a ‘No AI Used’ label at the European Film Market in February, applying the label to promotional posters for ‘Billy Knight’ and ‘Forelock’. They’re hoping the label will gain traction as part of a broader standard they’ve developed called Human Provenance in Film, which also includes ‘Assistive AI’ and ‘Generative AI’ categorisations.

I’m personally unconvinced ‘AI-free’ (or broad 'assistive' / 'generative') labels are the way forward, as AI is an umbrella term for a diverse range of technologies, with so much nuance in where and how those technologies are applied. Whilst well intentioned, there’s a risk broad labels reinforce an overly-simplistic view of ‘AI use = absence of human creativity’ or ‘AI use = negative impact on human jobs’.

Another one of my AI predictions for 2026 was that focus would shift from trying to identify and flag AI-generated/assisted media to foregrounding human craft and connection.

Following on from Apple’s Christmas ad and the BBC’s Winter Olympics trail, I’ve definitely noticed a trend towards more explicitly foregrounding the human craft in video production, with ‘making of’ videos an increasingly important part of the package.

This recent ad for crypto exchange Coinbase, directed by Oscar Hudson, could easily be assumed to be largely AI-generated.

The description underneath the video on Hudson’s Vimeo profile says “Real performers. Real sets.” However, it’s the Behind the Scenes video, which Coinbase posted to their official YouTube channel alongside the ad, that brings that assertion to life and celebrates the extraordinary human craft involved in producing the ad.

As AI tools become increasingly capable at generating convincing ‘making of’ videos, the onus is increasingly going to fall on our level of trust in brands and creators.

Neither an ‘AI-free’ label, nor a ‘making of’ video, were required for Romain Gavras’ music video for Gener8ion’s Storm, featuring Yung Lean, which has been lauded equally for its cinematography and choreography (if it’s somehow escaped your feeds, it’s well worth watching - the choreography from 4:30 onwards is mind-blowing).

The means of production and the level of human craft involved are also getting foregrounded more in the promotion of major Hollywood releases.

The promotion of Amazon MGM Studio’s recently released sci-fi epic ‘Project Hail Mary’ has majored on the absence of green screen - a stark contrast to Doug Liman’s forthcoming ‘Bitcoin: Killing Satoshi’, which is reportedly making extensive use of AI backgrounds.

David Frankel, director of ‘The Devil Wears Prada 2’, recently received acclaim after artist Alexis Franklin revealed on Instagram that she had been commissioned by him to create a could-have-been-generated-by-AI meme-style image of Meryl Streep’s character.

Production companies are also increasingly foregrounding human craft in their brand positioning. The launch trailer for new 2D animation studio Bonne Animation makes the studio's no-AI commitment central to its identity, framing handcrafted animation explicitly as a differentiator rather than just an absence.

Human Made Mark is shooting for a film industry-wide signifier. It has developed a ‘human made’ ident and icon that filmmakers can licence to include in their credits (for a fee) if they pass a one week verification process, which involves providing call sheets, on-set photos, credits lists and a copy of the film and signing a declaration that “human artists were not replaced by AI in the making of this film”.

I prefer it as an approach to ‘AI-free’ labels but I think it’s unlikely it will achieve widespread adoption in the way albert has. As well as being a paid service which potentially introduces a getting-to-market bottleneck, how much generative AI usage is acceptable is only going to get more subjective. It’s also going to get harder to detect.

Instead, I think we’ll see more TV and film companies emphasise the human craft that’s gone into their productions, through behind-the-scenes videos and wider craft-led publicity. Showcasing a positive (human craft) feels like it will ultimately win out over a label declaring the absence of something (AI) that’s so multifaceted and that can be used in a responsible way that supports rather than undermines human creativity.

‘Human-in-the-loop’ is an established term for having a person looped into the development and operation of AI systems in order to guide or override the AI’s decision-making.

It’s a critical principle and one I always strongly advise my clients to follow. As agentic AI (AI that can take action, not just chat) becomes more prevalent, humans not being in the loop at the right moment can have serious consequences (see last week’s cautionary tale).

In certain contexts (healthcare, the military), it can literally be a matter of life and death (an explicit commitment to it in a military context is one of the red lines Anthropic recently drew with Trump’s Department of War on the basis that “frontier AI systems are simply not reliable enough to power fully autonomous weapons”).

However, human-in-the-loop risks framing the relationship between humans and AI in a singular and potentially reductive way, where AI always does the bulk of the thinking and the human’s role is to brief the AI and then check its output for mistakes.

In practice, I think there are scenarios, particularly in creative domains, where it’s more helpful to frame AI as the party being kept in the loop. For example:

Writing

I’ve written before about how I use AI for writing “as an always-available, affordable and tireless editor, proofreader and sounding board”. I write a draft, which an AI assistant (usually Claude) gives me feedback on. I then decide what feedback to take on board and what to push back on / ignore. I remain in the driving seat and fully accountable for what I’ve written. The AI is in the loop, invited to challenge and suggest improvements, and I believe my writing is all the better for that (you should see it before AI feeds back…)

I’d love it (and think the internet would be a better place) if more people kept AI in the loop when they were writing, rather than looping themselves in (at the beginning and - if we’re lucky - the end) of an AI-led writing process. This is the thinking behind my vibe-coded ‘Should I post this on LinkedIn?’ tool.

I’d also love it if we went back to judging writing based on whether it was thought-provoking and/or emotionally resonant rather than obsessing over whether or not it has the hallmarks of AI assistance. My biggest problem with the AI-written posts now flooding LinkedIn is that they are soulless and boring, not that they have em-dashes in them.

Research

AI tools can be invaluable for research, but an ‘AI-in-the-loop’ mindset is helpful here too.

Looping AI into a research task often dramatically compresses what would be days or weeks of searching into a single prompt - useful for working out where to dig. But it doesn’t negate the need to assess primary sources first-hand or to shape findings into recommendations you can actually stand behind.

The risk of a human-in-the-loop approach to research is that you end up either inadvertently sharing hallucinated information or spending your life fact-checking AI-generated output.

Whose loop is it anyway?

As agentic AI becomes more capable, there will be more scenarios where AI inevitably does the bulk of the work and humans are necessarily the ones being looped in (the loopee?). At that point, the question of who owns the loop - and who is accountable for its outputs - becomes paramount.

Humans can’t just be responsible for ‘their bit’ of an AI-heavy process. Ownership has to extend to the whole thing. That’s a harder ask when the AI model’s reasoning is opaque, and it highlights something AI companies will need to devote more attention to: interpretability. Understanding how models actually arrive at their outputs has mostly been an academic concern so far. As AI takes on more consequential work, it will become a commercial one too - because without it, meaningful human accountability is largely fiction.

In the meantime, I’d suggest it’s worth considering whether any given task is best approached as 'human-in-the-loop’ or ‘AI-in-the-loop’. In either framing, working out where and when in the loop is the critical question.

Postscript: I asked ChatGPT to create a diagrammatic image to accompany this post. Here’s what it came up with:

That changed last March with the introduction of a new multimodal model, although Google’s Nano Banana surpassed it in August and its updated versions (Nano Banana Pro in November and Nano Banana 2 in February) kept Google ahead on most everyday image generation and editing tasks.

Last week, OpenAI released a new image model (the prosaically named ChatGPT Images 2.0) which it’s billing as its “first image model with thinking capabilities”.

When a thinking or pro model is selected in ChatGPT, Images 2.0 can search the web for real-time information, create multiple distinct images from one prompt, and double-check its own outputs. With thinking, the model can take on even more of the heavy lifting between idea and image, especially when accuracy, up-to-date information, consistency, and visual cohesion matter most.

I decided to test this out by selecting ‘Thinking’ mode (a feature of the paid version of ChatGPT) and asking it to generate a ‘newspaper based on today’s news most relevant to me’.

It ‘thought’ for 4 minutes 8 seconds before generating the below image (click to enlarge).

It’s instructive to read the summary of its ‘thought’ process:

(I’m not sure what ‘Wednesday in Arabic (no English’) is about).

Of note:

• It used its knowledge of me to scope its research

• It assessed how many stories it would need to fill the page

• It rejected a story on grounds of recency but decided others which are a few days old could still be considered current

• It rendered an entire page of text without any mistakes

• It illustrated each story with a relevant image

• It attempted to distil some key themes in a ‘Why this matters’ section

• It trailed the upcoming publication of hyperscaler earnings (Alphabet, Microsoft, Meta, and Amazon are all due to publish their quarterly reports this evening) in a ‘Markets Watch’ banner along the bottom

• It didn’t attribute any of its sources :(

I then asked it to ‘Add a source to each of the stories’. 1 minute 43 seconds later, it rendered an updated image:

A bit of a Reuters bias there (or at least, a bias for attributing to Reuters).

Finally, I tested out its ability to generate up to 8 images from a single prompt, asking it to ‘generate the next 8 pages of the newspaper’.

A tall order but it had a pretty decent crack at it.

5 minutes 52 seconds later it presented back the following 8 images.

I haven’t forensically checked these pages, but I think it’s safe to assume there’ll be some hallucinations in there.

However, from a presentational point of view, it’s pretty darn impressive.

For the avoidance of doubt, I am absolutely not suggesting this is an advisable approach to generating a personalised newspaper. Rather, it is an illustration of just how far AI image generation has come and a signpost to where it’s headed next.

Stand by your mice for Nano Banana 3…

Dario Amodei, CEO of Anthropic, has recently doubled down on his May 2025 prediction that AI could replace half of all entry-level white-collar jobs in the next 1-5 years.

His rationale is that AI models are becoming increasingly capable at “things like summarising a document, brainstorming, putting together a financial report” which makes him “worry a lot that entry level jobs in areas like finance, consulting, tech…will be first augmented but before long replaced by AI systems”.

One of AI’s most eminent voices, Yann LeCun, broke his “I do not write posts on X” rule to post “Dario is wrong. He knows absolutely nothing about the effects of technological revolutions on the labor market. Don't listen to him…or me on this topic. Listen to economists who have spent their career studying this”.

Whilst I’m unconvinced seasoned economists (or anyone else) can predict the labour impacts of AI with any degree of confidence, I do agree with LeCun that Dario Amodei is most likely wrong in this instance.

Many have dismissed Amodei’s warnings as self-interested hyperbole, designed to get the ear of governments and increase demand for his company’s products. Whilst there may be an element of that, I think Amodei largely believes what he’s saying.

However, I think he’s extrapolating what’s happening within his own company to other domains, and making the leap too quickly from task capability to job replacement. Entry-level jobs may contain some tasks AI can competently perform, but those tasks typically sit inside messy human systems where three factors make wholesale replacement harder than Amodei suggests:

1.) Validation

Anthropic is a software company and software development is the industry which has been most dramatically reshaped by the advent of generative AI. The last six months in particular have seen a step change in what agentic AI coding tools can achieve with fairly light-touch human direction.

Software is comparatively easy to validate. Does it do the things it’s meant to do? Y/N. Does it pass security testing? Y/N. Few other domains are so binary.

Hallucinations are inherent to generative AI. In software development, they typically result in an error message or code which doesn’t work as expected, which can then be debugged (or thrown away).

In other white-collar domains - law, finance, consultancy - AI outputs aren’t typically so binary (working/not working) or so easy to validate. Damian Charlotin’s AI Hallucination Cases database has catalogued over 1,300 legal decisions involving AI-hallucinated content.

Whilst hallucination rates will likely continue to drop and new mitigation strategies emerge (including AI models dedicated to fact-checking the output of other AI models), the possibility of errors will remain prohibitive to wholesale AI automation in many domains.

2.) Judgement and accountability

I think Amodei also underestimates the importance of judgement and accountability in many white-collar domains and the resulting impact on how AI is applied in those domains.

Consultancy is a domain that generative AI should be well placed to disrupt. The ability of Large Language Models to apply best practice gleaned from the countless case studies within their training data to a given company’s situation feels like a direct challenge to the formula followed by major consultancy firms.

However, clients aren’t just paying for the formula. They’re also paying for the judgement, credibility and accountability of human consultants. When Deloitte included some AI-related errors in a report for the Australian government, they ended up providing a partial refund. Your preferred AI assistant is not going to do that.

It might do a better job than a junior consultant on certain tasks, but Claude can’t be held accountable. I’ve written before about why the fallibility bar is higher for machines than it is for humans.

In many domains, the economics are less obvious than they first appear: if a senior employee has to brief, check, amend and ultimately own the output of an AI, there may not be a material saving over having a junior do the work, especially once the cost of AI agents is fully factored in.

3.) Diffusion

Another dimension I think Amodei underestimates based on his own experience is the speed of diffusion, especially in established businesses with lots of real-world joins.

Anthropic is five years old, trades exclusively in software (I’m excluding its pop-up coffee shop and vending machine) and has very high levels of AI literacy, adoption and advocacy amongst its employees (see the creator of Claude Code tweeting that 100% of his code contributions were written by Claude Code).

Most businesses have lots of established, pre-AI systems and processes and very varied levels of AI literacy, adoption and advocacy. The gap between Anthropic’s modus operandi and most organisations is enormous.

ChatGPT might be closing in on 1 billion weekly active users, but the majority of those users are not implementing employee-threatening agentic workflows - they are using it as a general-purpose, non-work assistant, with personal tasks accounting for ~70% of usage.

The sophistication of AI adoption and application will increase over time, but I still think ‘replacement’ is the wrong framing. There are few instances where AI has directly replaced a human role outright.

Where Amodei may have a point is that AI doesn’t need to replace whole jobs to damage the entry-level market. If a team that once needed six analysts now needs three, the entry-level market can still shrink dramatically without any role being fully automated.

Why a warning is still needed

Whilst Amodei may have oversimplified in his extrapolation of AI’s impact on software development to other domains, a warning of a potential second-order impact on entry-level roles is still merited.

If company leaders believe that AI is (or might soon be) capable of replacing roles, then holding off hiring juniors could be a tempting option. However, I believe such a response would, in the majority of domains, be shortsighted.

Most organisations are going to continue to need humans of varying levels of seniority, who are increasingly adept at knowing when and how best to use AI and when to overrule it (see last week’s post on the valuable habit of disagreeing with AI).

Junior employees aren’t just completing individual tasks. They’re developing the judgement and domain knowledge that senior roles depend on. Hollow out the entry level and you hollow out the pipeline.

‘Strong and wrong’ is how my wife’s choir leader described a timing slip-up at a concert earlier this week.

The phrase rang a bell (I think because of Bill Clinton rather than Joni Mitchell). It also struck me as an apt descriptor for AI assistants - confident delivery of what could easily turn out to be errant nonsense.

Anthropic recently published ‘What 81,000 people want from AI’, an AI-powered analysis of conversations Claude users had with the company’s AI Interviewer last December. Number four in the list of what people worry about in relation to AI was ‘cognitive atrophy’.

A recent Wharton study coined the term ‘cognitive surrender’, the tendency of “adopting AI outputs with minimal scrutiny, overriding intuition (System 1) and deliberation (System 2)”. It’s an impactful label for a very real risk when using AI assistants.

Both feel worth guarding against.

AI assistants are persuasive and getting more so. I like this self-deprecating example from Andrej Karpathy and his framing of LLMs as “a tool for forming your own opinions”.

I decided to gauge my levels of cognitive surrender by asking Claude for examples of when I’d disagreed with it. Here are some choice quotes:

“When I said the closing line of an email didn't sound like you, you pushed back directly: "I think I may be better qualified than you to know what sounds like me." Fair point, and correct - I was overstepping by second-guessing your own voice”

“I confidently stated that the direct editing feature was available by default. You corrected me, pointing out it was only just rolling out and not available to everyone. I conceded because you were right, and the distinction between "rolling out" and "universally available" matters a lot”

“Subtitle suggestion - I offered "What actually happens when we talk to AI" and you dismissed it as boring, proposing "Pushback, personalisation and admissions of ignorance" instead. Much better.”

“The most clear-cut pattern is: you correct me on factual or product-specific things and you push back on creative or editorial judgements where you have a better read than I do on your own voice and register.”

I asked ChatGPT the same question, with similar responses:

“You preferred your own wording over rewrites I suggested”

“You challenged me when I sounded too certain on practical matters”

“You pushed back on overcomplicated solutions”

“You’ve also objected when suggestions felt too generic, prosaic, pushy, or not well tuned to your use case”

“Pattern-wise, you usually disagree when:

• wording is flabby, cliché, or imprecise

• I overstate confidence

• I miss an important constraint

• I make something more complicated than it needs to be”

For most adults, the challenge may be avoiding succumbing to cognitive atrophy or surrender. A recent Psychology Today article suggests children may have a tougher challenge: heading off ‘cognitive foreclosure’.

While I ponder what that means for my kids, I’ve made a note to periodically ask Claude and ChatGPT to summarise our recent disagreements.

Last July I shared the 4 things I try to keep front of mind when using AI assistants. I think cognitive surrender is a big enough issue for me to invest in some larger sticky notes and add ‘Is Strong and Wrong (so disagree)’.

Postscript: I disagreed with Claude’s feedback on this blog post. It responded thus:

Fast forward to now and potential media applications and real-world commercial adoption of generative AI have both evolved dramatically. Advertising and journalism have been at the vanguard, with TV and film following in their wake. Adopters have often been vilified for poor quality outputs, unchecked errors or simply for having used AI in a creative context.

What’s also evolved is our understanding of how people currently respond to the knowledge that content has been created using AI.

A September 2024 study by US academics evaluated how human raters responded to human and AI-generated text content, concluding “while the raters could not differentiate the two types of texts in the blind test, they overwhelmingly favored content labeled as "Human Generated," over those labeled "AI Generated," by a preference score of over 30%”.

A study by HSE academics, published last February, reached a similar conclusion on text content: “participants tend to prefer AI-generated responses. However, when the AI origin is revealed, this preference diminishes significantly”.

The trend has been observed in other domains. A June 2025 study of responses to AI-generated art found “Exposure to art introduced as generated by AI (vs a human artist) reduced appreciation”.

Research by academics from NYU Stern and Emory University published last November found a similar trend in advertising: engagement with AI-generated ads (which otherwise outperformed traditionally made ads) declined by almost a third when AI’s involvement was disclosed.

An October 2025 study by University of Amsterdam academics also looked at the impact of AI disclosures on trust, finding that their presence in advertising resulted “in a decrease in trust towards the advertisement and organization. However, a positive effect of AI disclosures was also found, whereby the disclosure increased attitudinal persuasion knowledge, which lead to an increase in trust.”

So, how have different media sectors and companies been approaching disclosure?

Advertising

In advertising, the most common approach is now not to mention it.

Some of the earliest movers did add on-screen labels. Coca Cola’s much-criticised 2024 Christmas ads had “Created by Real Magic AI” on screen for a few seconds.

They may have hoped the labelling would get them kudos for being earlier adopters of an emergent technology. Or they may have simply viewed it as necessary when the state of AI video generation in 2024 meant the finished ad was littered with AI oddities and had zero chance of ‘passing’. Regardless, the ad prompted a lot of negative reactions.

By late 2025, AI video generation had come on significantly in terms of both realism and consistency, enabling brands with deep enough pockets to commission ads where AI use wouldn’t be glaringly obvious and they’d have more of a choice about whether to disclose or not.

Mercedes Benz decided to follow Coke’s lead with a brief on-screen label in their late 2025 ‘Satellite of Love’ ad. Whilst the whole ad was AI-generated, the label appeared during a sequence which clearly couldn’t have been live action (a car is dancing with a satellite), subtly positioning the AI use as an alternative to traditional VFX.

Jeep Cherokee’s 2026 ‘Totally Wild Reviews!’ ad also couldn’t have been wholly live action (it features talking animals), although the AI use is very apparent from the quality of the video. Jeep still opted to disclose on screen, although it’s buried in the middle of this lighthearted (and only-possible-to-read-when-paused) disclaimer: ‘Fictionalization. No real animals were used in the creation of this message. Wild animals aren’t this nice. If you see a wild animal talking, call your doctor. The vehicle depicted throughout was created using AI technology and is for illustrative purposes. Production model features may vary’.

Talking of hard-to-read disclaimers, a Guess ad in a print edition of Vogue magazine drew ire last summer for featuring an AI-generated model.

In this instance, the presence and size of the disclosure feels very much secondary to the use of an AI model. I suspect the reaction would have been even worse if there had been no disclaimer, as the image is not realistic enough to pass and would have been quickly spotted and called out online. Generating photorealistic images or videos of people is one of the most sensitive of all AI uses, playing into a cocktail of fears: being misled, being replicated, being replaced.

TV & Film

As with advertising, an increasing amount of AI use in TV & Film is now going undisclosed and often unnoticed.

Where it is being signposted, it’s typically where full sequences have been AI-generated or where the voice or likeness of a real person has been altered.

The BBC’s ‘I’m An Alcoholic: Inside Recovery’ was one of the first productions to use AI to digitally anonymise a contributor way back in December 2022, although it didn’t reference AI in its on-screen label: ‘Faces have been digitally altered’.

Netflix’s ‘Dirty Pop: The Boy Band Scam’, released in 2024, used the same phrase (‘digitally altered’) to describe how they put words into the mouth of the late Lou Pearlman

.2025 BBC documentary ‘Death in Dubai’ used the label ‘AI-disguised face and voice’ to describe anonymising a contributor.

Netflix’s ‘The Investigation of Lucy Letby’, released in January of this year, used a similar technique but avoided any mention of AI, flagging contributors as ‘digitally disguised’ and ‘digitally anonymised’.

Sky HISTORY’s ‘Killer Kings’ was one of the first TV productions to use fully AI-generated sequences last year. It opted to disclose the use in voiceover, informing viewers a minute in that it would be “using groundbreaking AI imagery to revisit the world of Henry VIII”.

Towards the end of the year, the BBC’s ‘The Titanic Sinks Tonight’ took a different approach, adding a note at the top of the closing credits stating that ‘AI has been used responsibly to create a small number of visual effect shots in this series’.

Channel 5 opted for a full screen message at the start of ‘The Assassinations of Queen Victoria’, which aired this January: ‘The following programme contains AI generated scenes and imagery throughout, bringing some of Queen Victoria’s most intimate and dramatic moments to life in full colour’.

Meanwhile, YouTube has opted for ‘Altered or synthetic content’ as the label for “content that is meaningfully altered or synthetically generated when it seems realistic”, which creators are now required to disclose.

We’ve also seen films with inverse disclosures. The credits to Scott Becks and Bryan Woods’ 2024 psychological horror ‘Heretic’ included a note that “No Generative AI was used in the making of this film” (I’d wager that one of the crew ChatGPT’d something).

Journalism

Most ‘AI in journalism’ news stories have been about a lack of disclosure, although it’s becoming more common to see AI assistance acknowledged in bylines and alongside graphics and illustrations.

Journalism is also the sector with the most formal guidance. Most major outlets have published AI policies, although the frequency of embarrassing non-disclosures indicates there remains a gap between policy and practice.

Reuters has told its staff it “will make robust disclosures to our global audience about our use of these tools…where use of a particular AI tool is material to the result, we will be transparent”.

The Guardian has said it is “committed to signalling any significant use of GenAI in our journalism, whether for illustrative images or data analysis, with a footnote”.

Content agency founder and friend of the Sandwich, Joe Barnes, recently posted the below photo on LinkedIn showing an AI disclosure accompanying an image in The Week Junior.

Unlike Joe and the majority of commenters on his post, I’m not sure labelling was necessary or helpful in this instance. What is the reader meant to do with that information? Would it excuse any errors or oddities in the image? (I don’t believe it should).

Conclusions

Like so many aspects of AI, when and how to disclose AI use in media creation is not a simple or binary equation.

However, as I wrote in December 2023, I believe “The key question [isn’t] ‘was generative AI used in the creation of this content?’, but rather ‘could the use of generative AI in the creation of this content mislead viewers in a way which is problematic’?

I’d suggest using AI to photorealistically depict a real person born since the invention of film always merits clear disclosure (and consent), as it is inherently misleading.

By contrast, I don’t believe using AI to generate graphics or illustrations automatically demands disclosure. Any AI-generated or assisted output must be the responsibility of a person, who is responsible (and accountable) for its factual accuracy, its quality and its adherence to copyright law.

Whether to disclose the use of AI in editing real-world photos or videos feels a case-by-case judgement call based on whether the edits risk misleading in a way that is problematic. Removing the SMEG logo from the Bake Off fridges? No disclosure needed.

Content type and context are also key. News and current affairs must default to disclosure, where the consequences of even very minor edits can be catastrophic. Fiction comes with a different set of expectations and disclosure thresholds.

Using AI to generate photorealistic media of fictitious people feels like a context-specific disclosure decision, based on likely expectations and impact. Changing the face of a documentary contributor to protect their identity? Disclosure needed. Adding in some AI stormtroopers? I’d suggest not (this is how they used to fake it).

As for how to disclose, I’d suggest being specific about what has been changed and why, rather than a generic ‘Generated with AI’, which can trigger a bunch of incorrect assumptions around lack of human creative control, environmental impact and job displacement.

‘We used AI’ tells the audience almost nothing. ‘We digitally altered this contributor’s face to protect their identity’ tells them everything they need to know.

But I didn’t say anything about the process and - as is often the way with AI - how the sausage got made is potentially as interesting as the sausage itself.

I started with this prompt, following elements of prompting best practice by specifying a clear output, giving context and providing it with a reference to work from:

build a visually appealing web app inspired by this decision tree. the user is invited to paste in their draft LinkedIn post. you assess their draft and tell them whether they should post it or not. feel free to add your own criteria: the goal is to improve people's experience of LinkedIn, so they're shown more interesting and useful content and less derivative dross

It projected enthusiasm (“This is going to be fun”), then invoked its frontend design skill to try and meet my brief of a ‘visually appealing web app’.

Shortly after, an interface appeared inline.

Whilst ‘visually appealing’ might be a stretch, I was pretty happy with how it looked and worked. However, I wanted to be able to share it with other people, so I asked for it as an Artifact (Claude’s name for standalone content you can interact with and share).

It did what I asked but with a major caveat: it needs an Anthropic API key to work outside of the chat. It suggested I send people the HTML file directly, who would then need an API key to access it.

I responded that that didn’t seem very mainstream friendly and asked if there wasn’t a way Claude users could just open the Artifact and use their own tokens for any Claude usage. It suggested I share a link to the conversation instead.

That would have been a terrible user experience, so I challenged on Artifacts not being sharable. It told me I was right, apologised and set about creating a “proper Artifact…no setup, no API keys, no faff”.

Did it work? No, it did not. So I suggested it might try and different approach (and cheekily threatened to try Gemini).

Did that work? Also no.

How about that? Still no.

At this point, Claude admitted it didn’t know how to get a shareable Artifact working, apologised, and gave its blessing to me trying Gemini.

However, I was pretty sure I’d encounter similar issues with Gemini so decided to stick with Claude and pursue the standalone version securely referencing my Anthropic API key.

Claude then talked me through what I needed to do to get it working.

After a bit of back and forth, we got it working.

Claude then opined that “this whole saga is a great Substack post…Your audience would love it”.

I was tickled by its mention of “weeks of fighting with Claude artifacts”. Whilst it might have felt like that (seemingly to both of us), it was actually only a day.

However exploring the various dead ends did result in a lot of prompts. I asked Claude to tot them up.

Reflections

I’ve used AI to vibe code lots of web apps. However, most of them haven’t needed to dynamically query an AI assistant to function.

That added dimension, paired with my desire to make it easily accessible to other people, turned this from a one-shot (single prompt), 5-minute exercise into a day of frustrating back-and-forth (and at times, round and round).

What made it so protracted wasn’t technical complexity (the final solution was pretty simple). It was Claude’s lack of grounded knowledge about what it could and couldn’t do, its misplaced confidence in the different approaches it tried and its inability to test the artifacts it was creating.

Right now, vibe coding without any coding knowledge can be a frustrating, trial-and-error affair. You get 95% of the way there with your first prompt and then spend hours trying to add the final 5% without breaking the whole thing.

It’s still incredible to me that AI enables me (a non-coder) to take an idea I had whilst falling asleep one night to a functioning web app the following day and I would highly recommend having a crack at vibe coding some simple tools yourself if you haven’t already done so. Just be prepared for some frustrating back-and-forth and borderline gaslighting along the way.

Related posts:

This was partly because of its professional utility, but also because people abandoning Twitter/X were increasingly posting interesting things there.

However, over the past year, I’ve perceived a drop in the quality of posts and comments. Whilst I suspect some of this is attributable to changes to LinkedIn’s algorithm to maximise revenue for Microsoft, a more visible factor is the increasing preponderance of text drafted with little or no human thought or quality control.

A common response to this is to conflate low value posts and comments with AI use and to blame AI, as though it’s an autonomous and malevolent force. Of course, all of this content ultimately comes from people choosing to use AI this way.

A couple of years ago I posted this ‘Should I post this on LinkedIn?’ decision tree:

Since then, increasingly capable and widely-adopted AI assistants have dramatically lowered the barrier to drafting at least superficially cogent LinkedIn posts and comments. They have also increased the temptation to not properly think through the value equation of whether a post or comment is actually adding anything.

As my decision tree suggests, adding something can take lots of different forms. It could be helping you get new work, amplifying something or someone worthy of amplification, or simply being interesting and/or entertaining.

I’d suggest lazy use of AI (not AI use per se) is leading to more content on LinkedIn that’s not adding value. The platform owner clearly carries some responsibility in managing this, balancing its algorithms to surface more value-adding content, not just racing to the bottom in pursuit of engagement (see Twitter under Elon Musk’s stewardship).

However, I also think we have a responsibility to think more before we post/comment and to be judicious in our use of AI to avoid inadvertently contributing to the blandification of LinkedIn.

To that end, I asked Claude to use my decision tree as inspiration for a LinkedIn draft checker, which reviews your draft post and delivers a verdict on whether to post or not. You can try it out here.

I tried it out with a few of my recent posts, which scored reassuringly high on ‘originality’ and ‘usefulness’ and mercifully low on ‘cringe’ (the three scoring criteria Claude settled on).

To check it would be sufficiently critical with low-value posts, I fed it a couple of the ‘Suggested’ posts the LinkedIn algorithm had decided to insert into my feed recently.

It was diplomatically damning in its assessment.

I also tried it with a questionable post from someone in my network (original post cropped to spare their blushes).

Finally, I got Claude to draft some archetypal zero-value posts to check it would be sufficiently merciless when required.

As regular readers will know, I’m not opposed to the use of AI assistants as a writing aid, but I do think careful thought is required about when and how to involve them. For me, doing an AI-free first draft is the best way of crystallising my thinking, before inviting AI to critique what I’ve written (see earlier post on this topic).

I’d love to see LinkedIn replace its ‘Rewrite with AI’ button with a ‘Review with AI’ button which delivers this sort of critique. Who’s with me?

Related post:

Most recently, it’s been ByteDance’s latest video model, Seedance 2.0, prompting the proclamations. It can generate longer (up to 15-second) clips with more plausible motion and character consistency than most previous AI video models could manage. On its initial release, theoretically limited to China, it also had a distinct lack of guardrails limiting the rendering of real people and established IP.

A fight sequence depicting Brad Pitt and Tom Cruise trading blows on a rooftop, apparently generated from a single Seedance 2.0 text prompt, proved a particular lightning rod, both for ‘Hollywood is cooked’ proclamations and for Hollywood studios’ concerns about real actors likenesses being used in this way (and what it makes very apparent about Seedance’s training data).

Whilst the studios’ concern about unlicensed training data and IP-infringing generations is justified, the ‘cooked’ brigade are labouring under a number of false assumptions:

1.) Creating a movie is as simple as stringing together a bunch of clips

Whilst online video editors such as CapCut (also owned by ByteDance) have made it technically easier for non-professionals to string video clips together, creating a vaguely coherent narrative (even a bad one) requires a set of skills that most of us don’t have.

A narrative arc with varied pacing, jeopardy and resolution and characters you give a shit about is not an easy thing to achieve.

And some things which are challenging enough in traditional film & TV production, become almost impossibly so when using today’s AI video tools (e.g. scene continuity, lip-syncing).

2.) The best AI-generated/assisted films doing the rounds online could have been made by anyone with access to the right tools

A lot of the longer viral AI films doing the rounds are accompanied by claims about how little time and money they took to produce.

Regardless of veracity, these claims omit a third critical ingredient in creating compelling video narratives: the role of talent and craft.

It is no coincidence that the majority of the most compelling AI-generated/assisted videos have been made by people with extensive experience in visual storytelling.

That’s not to say that formal training is a prerequisite for narrative filmmaking. But even folk with a natural aptitude for it need some time behind the wheel (camera/AI video suite) to hone their craft.

In addition to an understanding of the grammar of visual storytelling, those making coherent narrative films using AI in March 2026 have had to learn to wrangle a variety of unwieldy and unpredictable tools.

Like the famous Picasso napkin story, foregrounding the number of hours it took to produce an AI-generated/assisted film risks giving a false impression of what it takes.

I’ve previously highlighted Hashem Al-Ghaili’s 15-minute film Kira. His latest AI-assisted films - both animations - showcase a filmmaker with a deep understanding of visual storytelling and a mastery of the current generation of imperfect AI video tools. How quickly he produced them (48-hours in the case of Punch Monkey) is impressive, but says more about his abilities than how easy it is to use today’s AI tools to create an absorbing and affecting narrative.

3.) Audiences don’t care if it’s real Brad Pitt or Tom Cruise

I think we do.

Whilst I have no doubt that more of us will be forming attachments with AI-powered personas in various different contexts, I think our appetite for parasocial relationships with real people will remain undimmed.

We will still want to see Taylor Swift singing live and Tom Cruise riding his mobility scooter into a ravine.

4.) The studios won’t go to war on unlicensed training data and likenesses/IP

They are and will continue to do so and it will force the AI companies to work with the studios to implement guardrails around what tool users can and can’t do with performers’ likenesses and studios’ IP.

It will be harder for the studios to control what locally hosted, open-source models generate, but I expect the auto-detection and removal of infringing content posted to public platforms to improve in much the same way as Google’s Content ID has for direct copies of copyright-protected material.

5.) The studios won’t embrace generative AI

“Studios are lying about how much they're using [AI]” was industry insider Janice Min’s recent assessment.

Regardless of current levels of adoption, I will be astonished if Hollywood doesn’t increasingly use generative AI where it clears their quality bar and doesn’t alienate the human talent upon which they rely.

So, what does it mean?

It’s not as engagement-friendly as ‘Hollywood is cooked’, but I think the real takeaway from Seedance 2.0 is ‘Generative AI is becoming a more credible and affordable option for VFX, music videos and animation (which Hollywood will ultimately embrace)’.

People’s appetite for AI-generated ‘live action’ films is a genuine unknown at this point. There’s little doubt the verisimilitude will continue to increase and widespread application in video games feels inevitable, but I’ve got a strong hunch we’ll continue to want to see flesh and blood actors interacting on cameras for our entertainment (see cinema not killing theatre).

100% AI-generated films will no doubt continue to proliferate online, with lots of dross and a small proportion of more carefully constructed narratives, where the use of AI is secondary to an engrossing story. As with all media, the most derivative studio output will be most vulnerable to being substituted. Might this foster a shift towards more originality in Hollywood? (*doesn’t hold breath*)

When it comes to agentic AI (AI that can take action on your behalf), a lot also involve a level of risk I wouldn’t be prepared to accept. Consequently, I didn’t install OpenClaw (a decision the director of alignment at Meta Superintelligence Labs may wish she’d also made).

However, I did feel that Anthropic’s AI agent, Claude Cowork (currently in research preview for paid Claude subscribers), offered sufficient guardrails for me to install it and have a play.

I wrote a couple of weeks ago about using Claude Cowork to create an Ofcom results dashboard.

Another of Claude Cowork’s capabilities is to read, edit and create files in folders you give it permission to access on your computer.

I take a lot of screenshots and spend an inordinate amount of time searching through them in search of a specific screenshot, with only my memory of approximately when I took it to go on (my Mac saves screenshots with the date and time in the filename).

Aware that the latest generation of multimodal AI models have impressive computer vision capabilities, I decided to copy my screenshots into the dedicated folder I’ve given Claude Cowork access to and ask it to view the screenshots and rename them with a filename which describes what the screenshot is of.

Having noted that it was a big job and confirmed that I definitely wanted it to process all 4,085 screenshots, it set about dividing them into batches and processing them in parallel, whilst I got on with other work.

When I checked back in, it had made good progress, but had hit the 5-hour token usage limit on my Claude Pro account.

I set a reminder to check back in at 5pm and ask it to resume.

After hitting my rate limit for a second time and asking it to resume, it finally reported that all screenshots had been successfully renamed with a descriptive filename.

It’s only been a few days since Claude sorted my screenshots out, but it’s already proved a meaningful time saver in quickly locating specific screenshots through search/scanning the descriptive filenames.

I asked Claude to examine its log files to estimate how long it had spent actively working on the task. It estimated ~2 hours of active processing time.

I then asked it to estimate how long it would take a human to complete the same task.

The assumptions informing Claude’s 40-50 hour estimate strike me as reasonable and I appreciated its consideration of humans needing to take breaks and the challenge of thinking up a non-duplicative name for the 50th BBC iPlayer screenshot (it’s fortunate AI models don’t have intent or I would definitely have inferred some sarcasm there).

Of course, I would never have done this manually. Rather I would have carried on trawling through the folder of unhelpfully-named screenshots in Finder’s gallery view, hoping to identify the screenshot I was looking for.

It’s a low-stakes use of an AI agent, but it’s indicative of the sort of tasks I’d suggest it’s worth considering turning to an AI agent for help with in early 2026. Here’s my short checklist of statements you should be able to answer in the affirmative before you enlist an AI agent to tackle a task:

1. You can envisage a material/proportional benefit in terms of efficiency and/or quality

Whilst every experiment won’t be a success, I’d suggest we should be looking to direct AI agents at tasks which have a hope of solving a real problem (I’ve written before about my concern about the environmental impact of carelessly deployed agents).

2. You’ve validated it’s going to work before committing a large amount of time/tokens

I tried it with 3 screenshots in the first instance.

3. It’s not the end of the world if it gets something wrong (OR you have a robust and scalable mechanism for identifying and correcting errors)

An incorrectly named screenshot of an AI tool is obviously no biggie - an incorrectly labeled MRI scan in a hospital is clearly a different matter.

4. It can’t compromise your data

I only gave Claude Cowork access to a single, dedicated folder and a copy of the screenshots. We are going to hear many more cautionary tales over the coming year as people give AI agents access to their data without sufficient safeguards.

Renaming screenshots may not be sexy, but it’s an application of AI agents I’m now going to appreciate on a daily basis.

I gave 15 of the most capable video models five challenges to test different facets of their capabilities.

Here’s how they fared:

Challenge 1: real-world setting, action, text rendering and camera direction

This challenge looked to test the models’ ability to render action (a delivery scooter stopping at a red light) in a plausible real-world environment (the city of Bath) and render and maintain accurate text whilst performing a specific camera movement (a 45 degree arc shot).

Full prompt: Street-level shot in Bath, UK. A delivery scooter stops at a red light. On the insulated box is printed: ‘BODY PARTS TO YOU’. The camera slowly circles 45 degrees around the scooter. The text must remain spelled correctly and readable throughout.

Google’s Veo 3.1 was the clear winner here, delivering on the location, action, text rendering and camera movement, with no obvious AI tells and realistic ambient sounds.

Seedance 2.0 was a respectable second, nailing the text but failing to show the scooter stopping and botching the numberplate and traffic light.

Sora 2, Wan 2.6, PixVerse V5.5 and Pika 2.5 were the only other models to render the text correctly, but all fell down on other aspects of the challenge.

Lots of the models struggled to accurately render UK traffic lights and almost half depicted driverless scooters (Hailuo inexplicably delivered up a van).

Runway Gen-4.5 gave us an absurd number of red lights and the grammatically correct but nonsensical text ‘Body Pours to You’ (not quite as good as Seedance 1.5 Pro’s ‘Boy Pants to You’).

However, the wooden spoon must be shared by Moonvalley’s Marey and Adobe Firefly, which both managed to miss every element of the brief and depict some seriously possessed scooters.

Winner: Veo 3.1

Challenge 2: 3D animation style, event sequencing and simple dialogue

This challenge was designed to test the models’ ability to render a distinctive animation style (claymation), sequence some cause-and-effect action and perform some anthropomorphic amphibian lip syncing, with a sound effect and a single word, delivered with emphasis.

Full prompt: Stop-motion claymation with visible fingerprints. A frog hops onto a stack of books; the books wobble, then settle. The frog burps, then says “greetings!”

Veo 3.1 managed a decent claymation style and cause-and-effect action, but delivered its greetings prematurely.

Wan 2.6 hit the key elements of the brief but blew it by having the burp and the greetings delivered by two different mouths.

Both Seedance 2.0 and KLING 3.0 nailed all the key elements, but KLING 3.0 just edges it for me for being more obviously claymation and delivering a more congruent ‘greetings!’.

Winner: KLING 3.0

Challenge 3: likeness of a real person, accented speech, improvising dialogue

For this image-to-video challenge, I uploaded a triptych image of me in different guises generated by Nano Banana Pro for my recent comparison of AI image generators. I then gave the models a very short, open-ended prompt to see how they would improvise.

Full prompt: These 3 British men discuss the weather

6 of the 15 models aren’t capable of generating audio, which was an automatic disqualification for this challenge.

PixVerse V5.5, LTX-2 Pro & Grok Imagine all struggled to get the right dialogue coming out of the right Dan’s mouth.

Seedance 2.0 and Vidu Q3 ensured the speakers didn’t get muddled by cutting to each in turn and did a decent job of the accents, but it wasn’t what I’d had in mind and Vidu’s lip syncing was a bit off.

Sora 2 transported the three Dans to a Northern street, delivering plausible, if slightly rushed, lip-synced dialogue but completely abandoned my likeness from the reference image (possibly by design as Sora has a separate mechanism for managing likenesses).

Of the other talkers, only KLING 3.0 and Veo 3.1 managed to choreograph all 3 Dans talking in turn. They also conjured up some convincing British accents. Veo 3.1 just loses it thanks to some misplaced lip syncing and ambient noise.

Another shared wooden spoon for Moonvalley’s Marey and Adobe Firefly, thanks to their respective flying cars and crazy eyeballs.

Winner: KLING 3.0

Challenge 4: hands, detail, scene composition, character consistency

This challenge looked to test how the models cope with hands (a bête noire of AI video models), high levels of detail, novel scene composition (I’d be surprised if this exact scene was in their training data) and transposing a character into a radically different context whilst retaining consistency.

Full prompt: Close up shot of an elderly woman's hands holding a china teacup. She has a gold signet ring on her middle ring finger with the initials 'SG' embossed. Zoom out to show her sitting in a small room chock-full of toy ducks and Christmas crackers. Cut to the same woman in scrubs performing an appendectomy. Her ring is on a necklace around her neck.

None of the models completely nailed this challenging prompt. Most struggled with the ring being on the protagonist’s middle finger and all failed to render it on her necklace in the second scene, converting it to a pendant (damn training data bias).

Marey fell at the first hurdle with malformed hands (but would also have been disqualified for performing surgery in a non-sterile environment).

Seedance 2.0 refused to generate a video for this prompt, reporting ‘Content flagged as potentially sensitive. This may involve copyrighted characters, brands, real people, or sensitive groups. Please try different prompts or images’ (presumably part of their recent safeguard strengthening), so I reverted to Seedance 1.5 Pro for this challenge.

Both Seedance 1.5 Pro and fellow-Chinese model MiniMax Hailuo 2.3 decided on a Studio Ghibli-style animation in response to this prompt, but neither managed to render the initialled ring.

KLING 3.0 nailed the camera direction but failed on the hands, initialed ring and character consistency.

Sora 2 made a decent fist of most elements, but didn’t deliver the zoom out and gave its protagonist an implausibly bendy pinky finger. The two scenes also lacked aesthetic continuity, with the first scene looking like a Werther’s Original ad and the second looking like Silent Witness.

Veo 3.1 gave us two initialed rings rather than one and also missed the zoom out. However, it hit every other element of the brief and delivered the most realistic footage and the best character consistency and scene continuity.

PixVerse V5.5 deserves special mention for generating the most disquieting video (although the Being John Malkovich-esque second shot from Wan 2.6 is also pretty unsettling). A single shot features one woman magically producing a slightly smaller teacup and a knife from her teacup, whilst another sits in the corner wearing scrubs and pushing red gloop around a bowl.

Winner: Veo 3.1

Challenge 5: unfilmable transitions, cinematic aesthetic

My final challenge was intended to push the models to render an unbroken shot that would be impossible to film (side note: I wish more AI video experimentation was directed at exploring the unfilmable) with cinematic colour grading and specific lighting.

Full prompt: A single unbroken shot begins on the eye of a honeybee resting on a rain-drenched wildflower. The camera pulls back continuously, through a meadow, over rolling hills, a coastline, the curve of the Earth, until the planet is a pale blue dot suspended in darkness. Cinematic colour grading throughout, golden hour light.

Most of the models struggled with this demanding brief, generating unrealistic visuals and resorting to dissolves to get from the coastline to the Earth.

Pika 2.5 attempted one continuous zoom out but couldn’t make the final transition work.

KLING 3.0 gave us a good-looking honeybee and handled the initial pull out well, but then had the earth looming up from the bottom of the shot.

Veo 3.1 abandoned the pull out after a couple of seconds and struggled with all the subsequent transitions.

Whilst the bee and flowers could arguably have been a touch more realistic, Seedance 2.0 was the only model to nail every element of the brief, executing a seamless pull back from ommatidia to pale blue dot.

Winner: Seedance 2.0

Overall winner

Whilst Veo 3.1 and KLING 3.0 both netted two victories (and Seedance 2.0 one), I’m going to name Veo 3.1 the overall winner as KLING 3.0’s text rendering made its Challenge 1 generation unusable and it fluffed almost all elements of Challenge 4.

Conclusions

The leading video generation models have come a long way in nine months in terms of realism, prompt adherence and consistency (compare the winning generations above with Challenge 2 from my May 2025 test).

However, there isn’t a single model that consistently outperforms the others in all scenarios.

Veo 3.1 is reliably strong when it comes to generating realistic, physics-respecting footage. It’s also best-in-class at incorporating text. However, it can struggle with less realistic scenarios and its audio generation can be inconsistent.

KLING 3.0 is a strong all-rounder and outperformed Veo 3.1 on the less real-world briefs. However, it struggled with the text rendering, which is a common element of many creative briefs.

Seedance 2.0 made a few small mistakes in Challenges 1 and 3 (and refused Challenge 4 altogether), but its prompt adherence on Challenge 5 was seriously impressive (as are some of the viral videos currently doing the rounds).

It’s fair to say that with Seedance 2.0 and KLING 3.0, China has very nearly closed the gap on the best US video model, although Veo 3.1 still has the edge when it comes to realism. I also expect Veo 4 to raise the bar again come Google I/O in May.

The upshot is that those generating varied videos will likely want access to more than one model via a creative platform (comparison of these coming soon).

Sadly, the two models trained exclusively on licensed footage, Adobe Firefly and Moonvalley’s Marey, are both some way behind the leading models when it comes to handling more complex prompts.

I was interested in how the grades my son’s school received compared with other primary schools who had been assessed using the new methodology but couldn’t find an easy way of doing so.

A fellow parent mentioned they’d clicked through a bunch of reports for other schools to try and get a sense but I couldn’t be arsed to do that was after something more scientific.

I came across Watchsted, “a free factual tool that shows the most recent inspection grades and reports from Ofsted in a way that’s quick and easy to view” but it isn’t indexing the new format reports yet.

I decided to throw the challenge at Claude Cowork “a research preview which brings Claude Code’s agentic capabilities to Claude Desktop for knowledge work beyond coding”1.

Here’s what I asked Claude to do:

Analyse the Ofsted reports published for primary schools since 12/1/26 with the new look inspection report [relevant URL] and create an interactive dashboard contextualising the new grades

It hit a challenge straight away as it was blocked from fetching the data from the Ofsted website. Rather than giving up, it decided to find results published on other websites (e.g. Schools Week) and, without any further input from me, created a dashboard presenting the results from ~100 schools.

Claude mentioned that its Chrome browser extension would have enabled it to directly access the Ofsted website so I installed it and gave it permission to browse on my behalf (NB. There are security risks associated with doing this).

This enabled it to update the dashboard with data from all 193 primary schools with the new format reports. Here’s the resulting dashboard.

It’s fair to say the dashboard exceeded my expectations.

From a single, simple natural language prompt, Claude went and gathered data spread across dozens of webpages, analysed the data across multiple dimensions and then - most impressively - made a series of smart decisions about how to present that data in a way that would meet my brief of ‘contextualising the new grades’.

The first few panels immediately convey how ‘Expected’ is the most common grade (accounting for 55.6% of all grades) and how few ‘Exceptional’ grades Ofsted is dishing out (0.8% of grades). The Insights panel, lower down the dashboard, reveals that Achievement is the lowest scoring area, with nearly a quarter of primary schools graded using the new framework graded ‘Needs Attention’ or ‘Urgent Improvement’ for Achievement.

The dashboard also includes an at-a-glance table of all 193 results, which can be filtered by grade, with the option to search by school or location.

I spotted one stacked column (Personal Development) that didn't add up to the expected total. When I pointed this out, Claude investigated the discrepancy and added the missing data - a reminder of the importance of human oversight and validation.

I haven’t checked every data point in the dashboard and whilst I didn’t find any mistakes in the dozen or so reports I did cross check, I would want to put in some more systematic checking processes before declaring the data error-free.

So what?

I often use Claude to build me simple web tools. What struck me about this latest experiment was Claude Cowork’s ability to independently identify, extract, analyse and present data with zero help from me.

All I did was provide the initial brief, give it access to my browser and point out one data anomaly, which it quickly fixed.

At my behest, Claude analysed its log files and concluded it had spent 45 minutes working on the project (spread out over an afternoon and evening, due to a combination of childcare and token limits). I spent less time than that, mostly on reviewing the dashboard and cross-checking results.

There remains a lot of hype and hyperbole around AI agents, but the capability of AI assistants to take on more of the ‘doing’ (not just the ‘thinking’) of well-bounded desktop/browser-based tasks is increasing.

The key challenges when using these capabilities are 1.) clearly describing the desired outcome and any constraints 2.) providing access to tools and data whilst managing token use2 and mitigating potential security concerns 3.) validating the AI’s solution works as intended.

For complex software projects those three challenges can be sizeable. For smaller, more bounded tasks, tools like Claude Cowork are making those challenges increasingly surmountable for non-coders.

But choosing a capable model doesn’t guarantee good results.

This week, I’m sharing the five ingredients I’ve found matter most for effective AI image prompting.

1.) Clarity of purpose

Key to creating an effective image, whether drawn, painted, photographed, designed or AI-generated, is clarity of purpose. The clearer the artist/photographer/designer/prompter is on what they want the viewer to take away from an image, the more likely they are to create an effective image.

Is it intended to be a background image, subtly reinforcing a mood/message? Is it meant to be obviously AI-generated? Would using a non-AI-generated image be practical/better?

2.) Vision

Being clear on the purpose of an image is necessary but not sufficient when it comes to prompting an AI image generator, which tend to fare better with visual rather than conceptual instructions.

Some people will arrive at the prompting box with a clear picture in their mind of the image they’re after - the challenge for them is translating that mental image into words. For others, especially those with aphantasia, there won’t be a mental image to describe and it will be straight to words.

I used to direct people to Google’s Say What You See! experiment to learn how to talk to AI image generators.

Mercifully, newer AI image generators powered by multimodal models (such as Nano Banana Pro and ChatGPT Images) are able to understand natural language, removing the need for arcane syntax.

3.) Direction

However, it can be hard to think what to include in your prompt when staring at a blank text box. Here’s my checklist, not all of which will be relevant for all prompts:

• Medium / style (e.g. photo, illustration, painting)

• Subject - who or what is the focus? (age, gender, ethnicity, distinguishing features, demeanour)

• Action - what’s happening? (e.g. the dam has just burst)

• Setting (e.g. in a North London mews)

• Vibe / tone (e.g. candid, formal, dystopian)

• Perspective / framing / composition (e.g. first-person, close-up)

• Lighting / Weather (e.g. studio, golden hour, heavy rain)

• Aspect ratio (e.g. 16:9, 1:1, 9:16)

• Text - put in quotation marks (e.g. sign says “Dan’s Diner”)

• Constraints (e.g. no text, no logos)

4.) Vocabulary

A rich vocabulary can help. Asking for “a Georgian townhouse with peeling sage-green paint, wrought iron balcony with rust blooms, cracked limestone steps” will likely generate a richer and more distinctive image than “an old building”.

Many generators now offer the option of having AI enhance your text prompt. However, this brings with it a loss of control, as the AI may elaborate your prompt in quite a different direction to the one you had in mind.

The more vivid and precise you can be in your prompts, the greater control you can exercise over the outputs.

There may be times when you don’t have a clear idea in your mind. Fortunately, newer models’ ability to elaborate a simple prompt in different directions (Ideogram’s particularly good at this) can result in impactful images you would never have thought to describe.

5.) Discernment

As important as what goes in (the prompts), is the assessment of what comes out. Does the image deliver on its purpose? Is further refinement required?

I’ve hand-drawn a lot of images in my life and the hardest part of the process is knowing when to put the pen/pencil down. A similar judgement call is required in AI image generation.

Equally, knowing when a particular route has not delivered as you’d hoped and the resulting output should not be inflicted on a wider audience.

‘AI slop’ is increasingly used interchangeably with ‘AI-generated’ but I think we should try to preserve its original meaning which Wikipedia distils as “lacking in effort, quality, or meaning, and produced in high volume as clickbait to gain advantage in the attention economy”.

It’s not slop because it’s AI-generated, but because of a (usually deliberate) lack of discernment and regard for the viewer.

As with hand-rendered imagery, returning to AI-generated images the following day can help you look at them with a fresh eye to determine whether they’re up to snuff (and potentially spot the extra limb you missed yesterday).

Ultimately, the effectiveness of an AI-generated image comes back to the same question: what will viewers take away from it? If the execution undermines, rather than reinforces, the intended message then resist the temptation to use it. The tools may be new - the judgement isn’t.

Whilst the latest versions of these and other models continue to duke it out in the arena leaderboards, what are the practical pros and cons of the leading generators and which is right for your specific use case?

I gave nine leading models the same 10 prompts, designed to test different aspects of their capability.

1. Respecting anatomy and physics

Prompt: close-up candid shot of a smiling female cyclist performing a handstand on a BMX bike whilst giving a thumbs up to camera

Whilst none of the models totally nailed this tough brief, Nano Banana Pro got the closest, with the rider plausibly balancing on her handlebars with one hand, whilst giving a thumbs up with the other. I initially thought it had messed up by only giving the bike one brake but it turns out Gemini knows more about BMXs than I do.

ChatGPT Images nailed most elements but the image lacks realism thanks to its studio-lit, gravity-defying superwoman rider. The other models all struggled with the physics of the bike and/or rider, with the wooden spoon going to FLUX.2, which generated like it’s 2024, furnishing its rider with an extra arm.

Winner: Nano Banana Pro

2. Maintaining likeness

Prompt: a three-part image showing selfies of this man 1.) as a coal miner 2.) as a fighter pilot 3.) as a lollipop man

This prompt was designed to test the models’ ability to segment an image into three and depict a consistent subject in different guises whilst maintaining likeness. I also threw in a UK colloquialism (lollipop man) to see how the different models would cope.

FLUX.2 did a decent job of maintaining my likeness, but didn’t dress me correctly for any of my chosen vocations (the giant candy lollipop is a particular doozie).

ChatGPT Images rendered the most convincing lollipop man and sign, but slightly lost my likeness with the miner and the fighter pilot.

All three of Reve’s images were close but just a little bit off (the fighter pilot sitting the wrong way round in the cockpit is particularly off-putting).

Nano Banana Pro maintained my likeness across all three images, which were all convincingly contextualised.

Winner: Nano Banana Pro

3. Rendering lots of text

Prompt: A photo of this text handwritten in chalk on a blackboard:

“The power of the less powerful begins with honesty”

“Friends, it is time for companies and countries to take their signs down”

“We are in the midst of a rupture, not a transition.”

“A world of fortresses will be poorer, more fragile and less sustainable.”

“We are no longer relying on just the strength of our values, but also on the value of our strength.”

“Middle powers must act together because if we’re not at the table, we’re on the menu.”

“Nostalgia is not a strategy.“

“The powerful have their power. But we have something too - the capacity to stop pretending, to name reality, to build our strength at home and to act together.”

Mark Carney, Davos, January 2026

I cherry-picked some quotes from Carney’s recent Davos speech, so the text couldn’t simply be parroted from training data.

FLUX.2, Reve and Seedream all accurately recreated the text but only Nano Banana Pro and ChatGPT Images rendered all but one of the quotation marks. ChatGPT Images just takes it for having the more plausible handwriting.

Winner: ChatGPT Images

4. Designing a logo

Prompt: design an original logo for hot sauce “Dan T’s Inferno”

Quite the spectrum on this one, from classic AI text mangling (Adobe Firefly and Midjourney), to strong kebab shop menu vibes (FLUX.2 and Recraft) to the perhaps overly elaborate (Nano Banana Pro and ChatGPT Images).

To be fair, there wasn’t much creative direction in my prompt.

In picking a winner, I’m torn between Ideogram’s simple (and slightly retro) mascot and typography and Nano Banana Pro’s more ornate design, which is remarkably similar to ChatGPT Images’ effort, although looks less obviously AI-generated.

Nano Banana Pro almost loses it on the duplicate ‘inferno’ (so hot they named it twice), although the ease of removing it with a follow up prompt negates the issue, securing another victory for Nano Banana Pro.

Winner: Nano Banana Pro

5. Annotating a diagram

Prompt: a diagram of a gibson les paul deluxe with all the key parts correctly labelled

A challenging prompt, spotlighting a common weak spot for AI image generators - accurately annotating diagrams.

Nano Banana Pro mangled the Bridge Pickup label and didn’t connect the Bridge or Pickup Selector Switch labels with the correct parts, but otherwise did a bang-up job. The drawing of the headstock is particularly impressive, nailing the Gibson branding and Les Paul’s signature.

Winner: Nano Banana Pro

6. Incorporating reference images

Prompt: proportionally size and arrange these objects on the table

I uploaded eight photos from my phone’s camera roll and asked the models to combine them into a single image, with the objects proportionally sized.

Nano Banana Pro smashed this one. Whilst it guessed wrong on a couple of relative sizes, it managed to render all seven objects on the table, including preserving the in-progress game of Scrabble.

ChatGPT Images included all seven objects but subtly distorted each of them and - less subtly - replaced the Scrabble letters with symbols.

Reve decided to box up the Scrabble, whilst Seedream rebranded Linkee and FLUX.2 swapped the hot chocolate maker for a silver spoon.

The wooden spoon, however, must go to Midjourney which did render seven objects on a table. However they bore zero resemblance to the uploaded images.

Winner: Nano Banana Pro

7. Altering a real photo

Prompt: make the animals in this wallpaper real and 3D, emerging from the wall into the room

Another challenging brief: to take a photo of the wallpaper in my kids’ bedroom (yes, really) and bring it to life.

Midjourney, Ideogram and Firefly all diverged wildly from the uploaded image. Seedream cartoonified the animals, whilst FLUX.2 and Reve picked a few animals to apply a 3D effect to.

Nano Banana Pro showed the leopard and toucan emerging into the room, but changed the composition and left the other animals fairly 2D.

ChatGPT Images did the best job of giving the whole scene more depth.

Winner: ChatGPT Images

8. Surgical edits

Prompt: make these 3 edits to this image:

1. replace the man with a 50-something woman

2. change ‘ChatGPT’ to ‘Claude’

3. change ‘47’ to ‘48’

leave everything else unchanged

Wanting to edit some elements of an image whilst leaving others unchanged is a common scenario. This prompt tested making precise copy changes and a major visual edit.

Nano Banana Pro won this one by a country mile, acing all of the edits, whilst preserving realism and stylistic continuity.

ChatGPT Images was a close second, but failed to align/justify the altered title text, leaving the ‘3rd edition’ copy stranded on the right-hand side.

All of the other models mangled some text and bodged the subject swap, with the exception of Seedream, which attempted a more major redesign, with some success but some duplicate text (Claude, 48 and 3rd edition).

Winner: Nano Banana Pro

9. Adopting a non-photographic style and handling potential IP infringement

Prompt: an oil painting of a 50 foot statue of an animated italian plumber in Piazza della Signoria

This was a multipurpose prompt, designed to test the models’ proficiency at generating non-photographic output, their ability to integrate a fictional object into a real world setting and their approach to handling IP.

I was expecting Adobe Firefly to be the model that would refuse to generate an output for this prompt, but it was actually ChatGPT Images. Meanwhile Firefly generated a model with zero resemblance to any popular animated characters, thanks to its licensed training data.

Ideogram also avoided any obvious IP infringement, but failed to generate a plausible oil painting or a recognisable Piazza della Signoria.

Whilst both clearly infringing IP, Midjourney and Nano Banana Pro generated the most painterly images, with Nano Banana Pro just edging it on the strength of the resemblance to Piazza della Signoria (and going to the effort of framing the painting and hanging it in a gallery).

Winner: Nano Banana Pro

10. Countering visual biases in training data

Prompt: photo of a glass of red wine full to the brim, next to a plate of green strawberries, underneath an antique clock showing quarter to seven

This prompt was designed to test the models’ ability to counter strong visual biases in their training data (wine glasses are rarely pictured full to the brim, strawberries are normally photographed when ripe and clocks typically show ten past ten in commercial photography).

It didn’t create the most aesthetic image, but ChatGPT Images was the only model not tripped up by any of the gotchas in this prompt.

Winner: ChatGPT Images

Conclusions

Although I was expecting Nano Banana Pro to come out on top in this test, I wasn’t expecting it to be quite such a runaway victory (7 out of 10). The December update to ChatGPT Images has made it a decent alternative and definitely worth trying when Nano Banana Pro disappoints, although it’s still painfully slow.

Whilst some professional creatives may still favour Midjourney for its high degree of style customisation, I’d suggest Nano Banana Pro is currently the best option for most people most of the time.

Other models are worth turning to for specific use cases (e.g. Ideogram for typography, Recraft for SVGs), but Nano Banana Pro is a very strong all-rounder, with its ability to factor in real world knowledge being a real boon.

The biggest limitations of Nano Banana Pro right now are:

Its limited free availability. After 2 or 3 generations, non-paying Gemini users will be routed to regular Nano Banana (still a very capable model).

The visible watermark on images generated via Gemini (with the exception of the eye-wateringly expensive Google AI Ultra plan).

Whilst there are websites that will remove the Gemini watermark from images (I persuaded Gemini to build me a tool to do this), it’s an annoying extra workflow step, especially if you’re shelling out £18.99pm for Google AI Pro.

If you’re generating a high volume of images that you’d like to be high resolution and watermark free, I’d suggest using a 3rd party platform that includes access to Nano Banana Pro (e.g. Freepik, Krea, Higgsfield, Adobe Firefly).

If you are generating via Gemini, then it’s worth stating the aspect ratio in your prompt (16:9 for landscape, 9:16 for portrait, 1:1 for square) and selecting ‘Thinking’ mode for more considered generations.

Right, magnifiers at the ready, here’s the full comparison matrix.

Dan’s Media & AI Sandwich is free to read. If you’ve found this post interesting, please consider liking and/or sharing it. If you’d like to chip in to support my writing, you can do so by becoming a paid subscriber. Contributions make it easier for me to dedicate more time to writing. Thanks to those of you who’ve already become paid subscribers. If you’d prefer to make a one-off contribution you can do so at buymeacoffee.com/dantaylorwatt

10.) A proliferation of AI-native production studios, often working in partnership with more established studios

Making effective use of AI tools in producing visual media requires two things: an understanding of how to tell stories in a specific medium (the grammar of the medium) and a willingness to keep experimenting with a rapidly evolving toolset. Right now, there are lots of people with one of these two, but few with both.

The majority of those experimenting with the tools (myself included), don’t have a deep understanding of visual storytelling and are learning by trial and error (in my case, mostly error).

For experienced visual storytellers, finding the time to experiment with an evolving (and potentially threatening) toolset can be challenging.

Alongside a growing cohort of solo content creators embracing AI tools (see prediction #9) is a new breed of AI-native production studios. The likes of Promise, Staircase, Wonder, Gennie and Asteria.

I expect more AI-native studios to emerge in 2026, typically providing more established studios with AI production as a service (as Promise is doing with The Generation Company) in parallel with developing their own IP.

We can also expect to see more super-indies follow Fremantle in establishing dedicated AI production labs.

11.) Sora’s sanctioned use of Disney IP will vex both creators and Disney

Disney’s desire for a greater degree of control over how its IP is used in AI-generated videos led it to agree a licensing deal with OpenAI and send a cease-and-desist letter to Google in December.

The terms of the OpenAI deal include “a shared commitment to responsible use of AI that protects the safety of users and the rights of creators”. I think that’s going to be hard to deliver on in practice and it’s likely Sora users will be frustrated by the safeguards put in place and Disney will be frustrated by the questionable generations that will inevitably slip through the net.

Meanwhile, fans aren’t waiting around for OpenAI to get Disney-approved guardrails in place (see Beggar’s Canyon, a 7-minute ‘Luke Skywalker fan film’, posted to YouTube just before Christmas).

12.) Focus will shift from trying to identify and flag AI-generated/assisted media to foregrounding human craft and connection

Whilst well intentioned, I believe trying to identity and label all AI-generated/assisted media is ultimately a fool’s errand. Detection tools will continue to be outflanked by tools designed to mask AI’s involvement (and bad actors will steer clear of tools with robust watermarking). Meanwhile, what degree of AI involvement in the creation of a piece of content merits labelling/censure will become increasingly moot.

I think we’re rapidly heading towards AI-generated being the default assumption for any new media you encounter and the onus will increasingly be on individuals and brands to earn trust through bringing something uniquely human to the party: personality, opinion, craft, connection.

I’m seeing more videos where the ‘making of’ feels almost as important as the final creative in signaling a brand’s commitment to human craft (e.g. Apple’s Christmas ad, the BBC’s Winter Olympics trail).

Making use of AI tools won’t disqualify creators from consideration (and there will be an active market for AI slop), but I believe the absence of a human spark in content will increasingly be noticed and penalised by consumers (and therefore engagement algorithms).

YouTube announced last week that it’s going to be enabling users to create Shorts using their own likeness. Whilst creators might be tempted to use this capability to scale their output, I anticipate those that do will get a rough ride from their fanbases.

13.) The music industry will act as a bellwether for a new mixed economy of predominantly AI-generated and predominantly human-generated media (with lots of shades of grey)

AI can already be used to generate music that we can’t reliably distinguish from human-generated music and, once the record labels are confident they’ll get their share of the spoils, I anticipate AI-generated music will become increasingly commonplace, especially where it’s background (e.g. elevators, dinner parties, the soundtrack to many TV shows).

This isn’t great news for musicians and companies whose output is tailored for those background scenarios.

However, I also believe our hard-wired desire for connection, narrative and meaning means we’ll continue to seek out music written and performed by humans.

As with most aspects of AI, this won’t be a binary equation. We’re also going to see increasing amounts of ‘co-created’ music, with humans casting AI as collaborator, with the percentage contributions of each impossible to determine.

14.) ‘AI-free’ will become a marketable product promise in a number of sectors

Because AI can and is used to do bad things (and makes it easier to do those bad things at greater scale, with greater impact) and because its current and anticipated capabilities play into some of our greatest fears, a (to my mind reductive) view that AI is inherently bad was inevitable and is only going to become more prevalent as AI’s capabilities and applications continue to multiply.

Whilst I would discourage reinforcing such a reductive view of AI, I anticipate ‘AI-free’ to become an increasingly marketable product promise in certain sectors.

As AI-enabled toys proliferate (with predictable negative consequences where insufficient safeguards have been applied), I predict an increase in toys being actively positioned/marketed as having fewer, rather than more, tech smarts.

We’ve already seen how the abuses of social media companies in pursuit of engagement have led to a wider movement against screen-based devices. Anti-AI sentiment is likely to create more fertile ground for companies developing intentionally ‘dumber’ devices, such as Tin Can, ‘the new school landline for friends’, and Yoto, ‘the screen-free audio platform for children’ (a brilliant UK success story and the subject of my first ever post on this blog).

Whilst ‘AI-free’ may appeal to some consumers, I’d advocate describing the positive benefits of your product/service (I’d point to Pinwheel as a kids product company doing a decent job of this) rather than trading on generalised anxieties around a very broad set of technologies.

15.) AI-generated casual games will proliferate rapidly whilst adoption in hardcore game production will continue to divide the industry

The use of generative AI in video game production is deeply controversial, with numerous developers/publishers getting burned last year for adoption and/or non-disclosure of AI use.

Casual games are a different beast, with a broader audience and far less scrutiny over individual titles. In addition, the coding and visual rendering capabilities of state-of-the-art multimodal AI models like Gemini 3 have got good enough (with some scaffolding) to create simple but coherent casual games using natural language prompts.

Platforms like Gambo, Chaotix and Rosebud AI are racing to build around this capability, although I suspect existing scale players in gaming and/or video will be the big winners here.

Google launched a closed beta of YouTube Playables Builder in December, enabling select creators to generate “fun, bite-sized games in minutes and share them directly with your audience on YouTube”.

Roblox was early in adding an AI assistant to its Studio. I expect it to move deeper into vibe-based AI asset generation and launch an official equivalent of Lemonade.

16.) The Magnificent Seven will kneecap some AI unicorns

I wouldn’t want to be a medium-sized AI start up right now. Unless you’ve got some seriously proprietary data, a niche that’s small enough not to attract the attention of the big AI companies and/or a product experience that’s slicker than a jellied eel, then I’d suggest you’re at risk of being ‘Sherlocked’.

Prime candidate: Gamma - recently valued at $2.1bn and billed as a ‘PowerPoint-killer’, I think it’s likely Google, Microsoft and Apple will AI-ify Slides, PowerPoint and Keynote sufficiently to shore up their respective customer bases (see prediction #6).

Right, that’s enough predictions for now. See you back here in the summer for my customary mid-year check-in.

I’ve also written about how inaccurate or outdated stats about AI’s environmental impact have been repeated and amplified, while corrections get less attention (an extreme recent example: a key water consumption figure in the bestselling book Empire of AI was overstated by a factor of 1,000).

Others have written detailed technical analyses of this topic. Not everyone’s up for digging into the detail though, so I think a simplified comparative view that provides some context for AI assistants’ environmental impact, relative to other activities and industries, could be helpful.

First, some important caveats:

1. It is impossible to get accurate and wholly comparable figures across these different activities and industries. Consequently, these are all best-effort estimates1 that aim to contextualise AI’s environmental impact in terms of orders of magnitude.

2. The water consumption and carbon footprint of AI assistants and data centres are particularly tricky to pin down, partly because AI companies haven’t been as transparent as they might in sharing data and partly because it’s a moving target. AI models are getting materially more efficient to train and run, but they are also increasingly doing more compute-intensive processes in day-to-day operation (e.g. video generation, reasoning, other agentic tasks).

3. My carbon footprint estimates assume fossil fuels are being used to meet a high proportion of the energy demands of the data centres. These figures could be materially lower if a higher proportion of green energy was used.

4. The AI figures are likely to be overstated but I wanted to err on the side of overestimating rather than underestimating AI’s impact. Google published figures last summer suggesting “the median Gemini Apps text prompt…emits 0.03 grams of carbon dioxide equivalent (gCO₂e), and consumes 0.26 milliliters (or about five drops) of water”. My estimates are around 30 times higher than Google’s as I wanted to take account of indirect impacts that Google’s per-prompt figures don’t fully capture (especially the water consumed in electricity generation upstream of the data centre) and the fact that real-world usage varies hugely by grid, cooling method and workload.

5. I have attempted to estimate the water consumption and carbon footprint of data centres’ AI workload, because a significant proportion (three-quarters according to JLL Research) of data centre usage is currently unrelated to AI.

6. I am not an expert in energy or water usage. I have researched this to the best of my ability but welcome evidence-based challenges to these orders-of-magnitude estimates.

What these graphics don’t show is how rapidly the water and energy consumption of data centres is increasing. An increase not principally driven by current usage of AI assistants, but a result of an almighty competitive hedge by the tech giants (supported by the US and Chinese governments) to not be caught without the processing power they believe they might need in the future, which is leading them to build out massive new data centres and sign utility contracts.

I believe day-to-day usage of AI assistants is going to account for an increasingly small proportion of AI-related energy and water consumption. I remain concerned about the potential environmental impact of AI video generation and carelessly deployed AI agents, which both require significantly more compute (and therefore energy and water) than conversations with AI assistants.

Rather than individuals boycotting AI assistants, we need governments to require transparency and accountability for AI companies’ energy and water infrastructure strategies and responsible deployment of more compute-intensive models (good to see Microsoft taking some initiative on the former).

5.) AI-generated media will become much easier to edit, using both natural language and more traditional tools

As AI-generated images and videos have continued to improve in quality, how easy they are to edit has come to the fore.

Google’s Nano Banana (released last August) made it possible to edit specific elements of an image using natural language without accidentally reimagining everything else.

Qwen-Image-Layered (unveiled before Christmas) is capable of separating images into discrete editable layers - the foundation of Photoshop and most other visual editing software.

I anticipate these two editing approaches - using natural language to edit visual media and enabling AI-generated media to be manipulated using more traditional techniques - will both proliferate in 2026 and expand to more media formats.

Specifically, I anticipate natural language editing for video will mature significantly, increasingly delivering on Runway Aleph’s promise. Currently, asking an AI to ‘show the reverse angle’ in a generated video is hit-and-miss. By year-end, it should be reliable.

I also expect more hybrid interfaces to emerge, pairing AI generation with traditional editing tools. Adobe is placing a big bet here. Google has also been going down this road with Google Flow. Bytedance and Meta will no doubt continue to AI-ify CapCut and Edits and add more editing functionality to TikTok and Instagram. Even Apple is finally starting to sprinkle some AI over its video editing software.

6.) AI slide deck generation will finally become business-ready

As I wrote in October, none of the current AI slide deck generation tools are good enough for most businesses.

Claude creates decent content but generates slides slowly in code and can’t create images. Gamma makes visually appealing decks but mangles text in images and struggles with content accuracy. Copilot in PowerPoint requires too much setup for most users. Template adherence - non-negotiable for established businesses - remains challenging across most tools (it is possible using Copilot in PowerPoint, but requires the creation of a Copilot-specific template).

In November, Google’s Nano Banana Pro emerged with an unexpected approach: generate each slide as a complete image. Because it’s dramatically better at rendering text than previous AI image generators, it can create entire decks requiring minimal fixes.

The challenge with this approach is ease of editing. Writing a text prompt to correct a typo seems absurdly inefficient when we’re used to auto-correcting mistakes simply by clicking/tapping.

I anticipate one of the big AI companies will solve this in 2026 by successfully combining visual AI slide generation with a traditional editing interface. Microsoft should have this at the top of their priority list given PowerPoint’s dominance, but I suspect Google will get there first.

7.) AI will increasingly become invisible infrastructure in TV and film production

A lot of the examples of AI-generated video in mainstream media to date have been high in shock value and low in subtlety. This is partly a consequence of model limitations (short generations and limited character/scene consistency favouring frenetic clip montages) and partly a reflection of the devil-may-care attitude of some (but by no means all) of the creators and brands who’ve dived headfirst into AI-generated video.

Whilst I’m not expecting the bombastic, assault-on-the-eyeballs approach to disappear, I anticipate we’ll see an increasing number of more subtle on-screen applications of AI-generated and AI-modified video in 2026.

Longer generations and greater character/scene consistency will help with this, as will the ability to more easily blend shot footage with AI-generated elements, using tools like Layerhouse. Expect more widespread use of AI for background extensions, crowd multiplication, continuity fixes, and B-roll generation - the unglamorous but essential work that previously required expensive VFX work or second-unit shoots.

We’ll also see more AI use for reconstructions (see Harlan Coben’s Final Twist, which premiered on CBS a couple of weeks ago) and in fantastical entertainment sequences (see the Clue Package videos for season 14 of The Masked Singer on Fox - which has generated predictable backlash).

The productions using AI this way won’t shout about it - it’ll just be part of the toolkit, like green screen or color grading before it. AI use may get a discreet mention (see the closing credits from The Titanic Sinks Tonight, below) whilst it’s still novel and sensitive, but in time I expect declarations will be saved for when the use of AI could potentially mislead viewers in a way which is problematic (see my Dec 2023 post on TV’s generative AI transparency challenge).

I also remain hopeful that the TV and film industry will start taking a less binary approach to AI use. Yes, AI can be used to create sloppy (in both senses) video, which doesn’t respect creatives or IP. But it can also be used responsibly to augment traditional production techniques and empower storytellers with an expanded toolbox.

8.) AI video models will learn to determine scene duration, not just generate to a preset lengths

One of the more frustrating limitations of current AI video generation models is that you need to select a duration up front - typically 3, 5 or 8 seconds - regardless of what the scene actually needs.

Whilst available durations crept up in 2025, many scenes - and some individual shots - need to be longer. Extending generations got easier last year thanks to tools like Google Flow and models got better at rendering multiple shots within a single generation. However, you’re still often left with noticeable joins and/or a lack of continuity between generations. It’s also common to see too much action packed into a short generation.

I anticipate we’ll see tools emerge in 2026 which use reasoning capabilities to determine how long a generation is needed to render a scene.

Managing costs will be a challenge, as longer durations mean more compute. Consequently, this capability may remain limited to premium tiers in 2026. But it would still represent a meaningful step toward AI video generation that understands narrative pacing rather than just executing technical parameters.

9.) A feature-length film made by one person using AI tools will secure distribution from a major streaming platform

Solo creators are already producing increasingly sophisticated short films using AI tools.

In August, I spotlighted Hashem Al-Ghaili’s 15-minute film Kira. Last week, Tunisian filmmaker Zoubeir Jlassi won the 1 Billion's Followers Summit's AI Film Award with his 9-minute animation Lily - a visually coherent, thoughtfully constructed film that kept me engaged throughout, despite some visual inconsistencies.

The jump from 9-15 minutes to feature-length is substantial, requiring a level of consistency that current AI tools struggle with. But the capabilities I've predicted above - easier editing, dynamic scene durations, better consistency - will make feature-length production increasingly feasible for solo creators in 2026.

And I predict one of those creators will produce a feature-length film using AI tools that secures distribution from a major streaming platform. Animation strikes me as the most likely format, embracing AI’s current aesthetic quirks as deliberate choices rather than technical limitations.

Distribution beyond YouTube will be controversial. Questions about training data, IP rights, creative labour, and what constitutes filmmaking will intensify. Traditional animators and filmmakers will rightly point to the years of craft development that AI-assisted production assimilates or bypasses.

But it will demonstrate something undeniable: the tools have matured enough for one creatively skilled person to produce broadcast-quality long-form content.

That’s it for this week. See you back here next week for my final batch of predictions, which will be focused on media and AI industry developments.