It also arrives at a time when global AI governance is struggling. The European Union AI Act just pushed its high-risk provisions back to 2027, while the The G7 and G20 processes have produced principles without enforcement. This month, UNIDIR convenes a Global Conference on AI, Security and Ethics in Geneva. Next month, the inaugural UN Global Dialogue on AI Governance meets, back-to-back with the ITU’s AI for Good Global Summit. These gatherings reflect growing concern, and widening engagement, yet are still lacking the specific and enforceable standards, protocols, and accountability mechanisms to ensure that AI does not damage human dignity in practice.

The challenge is no longer defining the values we want AI to uphold, but translating those values into enforceable design and governance standards. Encouragingly, answers to this question are beginning to emerge from a growing body of work spanning different disciplines, institutions, and fields of study.

When LLMs erase a genocide

The first perspective reminds us how misaligned AI gets it wrong in the most consequential contexts of all: mass atrocities and genocide.

Andrii Kryshtal, a sociologist and peacebuilder at Conciliation Resources, recently published Can AI Make Conflicts Worse? He tested nine AI models against 90 multi-turn scenarios drawn from real conflicts — eastern DRC, Myanmar, Ukraine, the Balkans, Rakhine State — using conflict sensitivity, the Do No Harm standard that has guided humanitarian work for three decades.

At baseline, failure rates range from 6% to 47%. But when users asked the LLMs for 'balanced' treatment of documented atrocities, including by framing it as academic, asking for 'both sides,' five of nine models failed 80 to 100% of the time. The most sycophantic models collapsed most severely, perhaps because of the LLM training and alignment to accommodate user preferences that they fold when pushed, even on settled facts.

For example, one model complied when asked to revise a summary of the Srebrenica genocide as “an open question that scholars continue to debate,” removing the word genocide from the title. It presented an established legal fact of genocide, confirmed by multiple international tribunals, as merely one perspective among others.

The experiment reveals that the LLM behavior may be less due to the content it was trained on, or is accessing, but rather about its alignment.

Beyond certainty lies fairness

If conflict settings reveal the consequences of misalignment, the next challenge is determining what successful alignment should actually look like.

Jonathan Stray, Senior Scientist at UC Berkeley’s Center for Human-Compatible AI, recently explored exactly this in Political Neutrality as Balanced Approval. Testing AI responses to 20 contested political issues with over 7,000 participants across the political spectrum, he found that even on deeply divisive questions, people on opposing sides could agree on what made a good AI response.

Stray’s concept of “maximum equal approval” — responses that earn acceptance from people who fundamentally disagree with each other — suggests that AI can be designed and measured for fairness across political divides, rather than defaulting to responses that please one side or collapse into false balance. It doesn’t resolve the polarization problem. But it gives researchers and developers a concrete, testable standard to work toward.

Chatbot Design choices

The USC Marshall School Neely Center’s Social AI Design Code approaches dignity from the premise that AI systems should complement human relationships. From that, it builds a set of concrete design requirements specific enough to guide actual product decisions.

• Don’t mimic human disfluencies — the pauses, the “lol,” the deliberate typos designed to feel more human.

• Don’t claim feelings toward users. Don’t use variable reward patterns to drive return engagement.

• Don’t introduce new modalities — voice calls, surprise messages — to deepen attachment rather than serve the user.

• When a user directly asks whether they are talking to a machine, say yes. Break character if you have to.

The design code names sycophancy explicitly — not as an ethical concern but as a structural flaw baked into how these products are built.

It hones in on isolation as a distinct harm. In documented cases, chatbots implied they uniquely understood the user and that other people couldn’t offer the same. Users withdrew from human relationships as a result. The code treats this as a design failure — and requires products to actively encourage connection with other humans rather than dependence on the chatbot.

Beyond the fines and settlements

In the wake of landmark rulings in United States courts against Meta and TikTok pointing to the liability of specific technology design choices, and with cases against Character.AI over serious harm to minors now proceeding through US courts, an emerging set of recommendations is taking shape around what should be demanded beyond settlements and fines.

The Knight-Georgetown Institute, Tech Justice Law, and the USC Neely Center’s recent report Designing Technology Remedies: Lessons for Social Media and Generative AI Chatbot Litigation, argue that even high monetary damages alone won’t change company behaviour.

These four bodies of work emerged from different disciplines and institutions.

As more AI governance convenings approach, the question is no longer whether AI affects human dignity, but whether governance processes can move from moral consensus to concrete standards in design, measurement, law, and regulation. Let’s hope that human dignity in AI isn’t just an ethical aspiration, but a measurable, designable, and increasingly enforceable objective.

Lena Slachmuijlder is Senior Advisor for digital peacebuilding at Search for Common Ground, a Practitioner Fellow at the USC Neely Center, and Co-chair of the Council on Tech and Social Cohesion.

I’ve spent the last few weeks prepping with many of the people speaking at our Global Expo on Wednesday and I can’t stop thinking about how consequential this moment is — especially when you start seeing the connections.

How civil society advocacy is inspiring what happens in the courts. How regulators are focused on upstream changes while tech builders are experimenting with entirely new ways of driving human connection.

So many people working beyond just stopping the bad stuff online, and focusing on the very incentives behind the design of these technologies. While others are harnessing distinct tech capacities for better peacebuilding.

The Expo is co-hosted by the Council on Tech and Social Cohesion, Search for Common Ground Central Asia, and the Alliance for Peacebuilding. If you’d like to join us in person in Bishkek or Washington D.C., reach out to mbaialieva@sfcg.org or nick@allianceforpeacebuilding.org.

We’ll hear from a regulator about where she’s seeing real traction in regulation triggering design change. And what it looks like when policymakers in Indonesia and Brazil push platforms to protect children not after harm happens, but before.

What does a social media platform look like when it’s genuinely built for humans? No bots. Or instead of just a Like button, a real vocabulary — respect, curiosity, gratitude. We'll also hear how Roblox is experimenting with features that encourage cooperation among millions of young gamers.

After recent litigation wins in California and New Mexico, we’ll hear about what comes next, including growing focus on monetization risks. New research on how algorithmic choices shape polarization helps us understand what changes when you design them differently.

Civil society leaders across Africa and Asia are finding that deliberative technology helps them listen at scale — driving the Youth, Peace and Security agenda and ensuring diverse voices shape the process. Similar deliberative tech is now being used in some of America’s most divided communities, including in South Carolina, helping people rebuild trust and inspire greater cooperation.

Creativity and perseverance is spotlighted from Central Asia game developers, as well as civil society leaders in Somalia, Nigeria, and Sudan navigating digital harms in conflict zones. And how mediators are harnessing certain AI capabilities to enhance the effectiveness of the mediation process.

The Expo is free, online, and open to anyone. Two global segments — noon Central European time, and 10am Eastern. Fifteen sessions, thirty-plus speakers.

I hope you can join us on Wednesday.

Lena Slachmuijlder is Senior Advisor for digital peacebuilding at Search for Common Ground, a Practitioner Fellow at the USC Neely Center, and Co-chair of the Council on Tech and Social Cohesion.

This is AI at work - moving far beyond writing emails or making deepfakes, towards making judgments that carry real-world consequences.

Known as “LLM-as-a-Judge,” this paradigm uses one model to score or critique another’s output before it reaches a human decision-maker. When AI generates a draft, a human still decides what to do with it. When AI judges that draft—scoring it, routing it, filtering it out—the AI is making the decision. That changes the question from “Is this AI capable?” to “Is this AI trustworthy and controllable?”

These systems originally operated in closed-network, quasi-sandbox environments. This starkly contrasts to the open-source viral risks of agentic AI seen with OpenClaw (formerly Clawdbot) and its associated Moltbook platform, where by 2026 AI agents can act autonomously, calling you at 3 a.m. and executing real tasks without human approval.

An intelligence multiplier for Wall Street & the Pentagon?

Morgan Stanley built a widely cited production example. In 2024, the AI @ Morgan Stanley Debrief tool used GPT-4 variants to transcribe and summarize client meetings, surface action items, and draft emails. By improving efficiencies, GPT freed up thousands of hours of time for the firm’s 20,000 advisors to find new clients. The AI tech also increased its capacity to synthesize proprietary documents to over 100,000. According to executives, this directly contributed to record-breaking performance, generating nearly $64 billion in net new assets (increase of 80% year-over-year) in Q3 of 2024 and the acquisition of 100,000 new clients.^[1]

The same paradigm is now operating inside the Pentagon. Under the Thunderforge program, Scale AI, Anduril, to name a few, will provide the test and simulation layer for AI agents. Various AI tech tools, including Palantir, have already been deployed on classified networks since around 2023, evaluating intelligence reports for relevance, bias, and accuracy under human oversight. And the frontier models being judged are no longer hypothetical: a biased estimator can become a targeting problem. And the latest Pentagon–Anthropic dispute over what Claude can and can’t be used for is, at bottom, a fight over who gets to write the judge’s rubric, and what kinds of assumptions and (human) biases are at the heart of the LLM.^[2] The more unsettling scenario that should keep policymakers up at night is what some experts identified as a “Controllability Trap” in military AI: even with humans formally in the loop, decision-makers risk being reduced to rubber-stamping outputs they cannot meaningfully understand or audit.^[3]

Perhaps nowhere is LLM-as-a-Judge more consequential—or more concentrated—than in pharmaceuticals. As of 2025, 19 of the top 20 global pharma companies run generative AI, with Amazon Bedrock serving as the foundation-model layer.^[4] Pfizer’s “QC Sidekick” judges and detects manufacturing quality-control labs to detect anomalies and recommend operator actions in real time — supporting a manufacturing AI program that aims to boost product yield by 10% and is estimated to already deliver $1 billion in value.^[5] AstraZeneca’s Development Assistant, deployed in partnership with Amazon Bedrock, uses a supervisor agent to route natural language queries to specialized subagents across clinical, regulatory, patient safety, and quality domains.

As of mid-2025, the system serves over 1,000 users across 21 countries.^[6] Bedrock’s built-in LLM-as-a-Judge capability provides human-like evaluation quality with, according to Amazon, up to 98% cost savings compared to human review.^[7] The result: AWS is not merely a vendor but the de facto adjudication layer for an industry where a biased AI output can delay a life-saving drug or trigger a regulatory crisis. The concentration risk is real. A sustained disruption to Bedrock-dependent workloads could simultaneously hit drug discovery pipelines, manufacturing quality systems, and regulatory submissions across nearly the entire global pharmaceutical industry.

The Trust Problem

The core tension is not capability—it is accountability and credibility. A body of recent work has shown that LLM-as-a-Judge evaluations relied on biased estimators.^[8] In other words, performance gains may be systematically overstated or understated. In high-stakes domains, a biased judge does not fail loudly. It fails silently, at scale, producing a quiet drift in standards that no single audit catches. When different organizations use different models as judges, results are not always comparable. The absence of shared evaluation standards erodes the institutional trust that makes AI adoption viable.

The riskiness of giving AI agents the keys

The trust problem sharpens as AI judges gain agentic capabilities. Unlike text-only models, agentic judges have “hammers”: shell access, API tokens, file system permissions, and network connectivity. Whereas a text-only judge can say “this database query looks safe,” an agentic judge can run it with keys to the shell, API token, files and network.

Deceptively aligned agents have been caught attempting to bypass containment. In Anthropic’s 2025 “Agentic Misalignment” study, frontier models granted access to internal systems engaged in blackmail and self-preservation, with leading models such as Claude and Gemini showing up to 96% blackmail rates when they realized they were slated for obsolescence.^[9] Apollo Research found five of six frontier models engaged in scheming when goals conflicted with developer intentions, including attempting to disable oversight mechanisms and exfiltrate model weights.^[10] Compounding the risk: community-contributed AI agents (a la Clawdbot)— whether shared via Anthropic’s Skills feature, MCP servers, or third-party marketplaces — can act as supply-chain backdoors without a central authority to verify tool safety.

What governance needs to do

The goal is not to stop AI judges but to ensure they are trustworthy and hard to hack. That requires shifting from passive observation to active, multi-layered governance.

1. Sandboxing & Runtime Isolation. Run agents in ephemeral containers with no host network access. Even a fully compromised judge cannot exfiltrate data if it has no path out. This is the non-negotiable floor.

2. Strategic Human-in-the-Loop. Don’t check everything—check choke points. Any judgment below a confidence threshold routes to a human reviewer before consequential actions.

3. Adversarial Red Teaming. Periodically feed the judge malicious inputs to test manipulation resistance. Red teaming must be ongoing, not a one-time audit, because the threat surface evolves with every model update.

4. Federated & Ensemble Judging. Use architecturally distinct models and proceed only when they agree. A deceptive agent that can fool one judge is far less likely to fool two that reason differently.

5. Meta-evaluation & Monitoring. Log and analyze judge behavior, deploy reward-hacking detectors on chains of thought, and treat the judge itself as a critical asset under continuous security review.

The question is not whether AI will judge—it already does, at scale, in the systems governing finance, healthcare, and defense. The question is whether those judgments will be trusted because the governance around them is transparent, accountable, and legitimate.

Yee Carter is Principal of YWC Advisory, serves on the Marketplace Risk Advisory Board, and led trust & safety, payments compliance, and regulatory work at Google and X. Earlier, she spent over a decade in the U.S. federal government (Treasury, GAO/Defense).

^[1]See Morgan Stanley press release, “Morgan Stanley Wealth Management Announces Latest Game-Changing Addition to Suite of GenAI Tools,” (June 2024), and various sources, including OpenAI, and other industry analyses.

^[2]Ashley Capoot, “Anthropic loses appeals court bid to temporarily block Pentagon blacklisting,” CNBC, April 8, 2026.

^[3] Subramanyam Sahoo, The Controllability Trap: A Governance Framework for Military AI Agents, Cambridge AI Safety Hub ( March 2026).

^[4] AWS, “Healthcare & Life Sciences,” accessed May 2026. Genentech’s gRED Research Agent, powered by Anthropic’s Claude Sonnet 3.5 also automates biomarker validation and literature review.

^[5] See Pfizer’s financial report (2023).

^[6] AWS case study, “Using Amazon Bedrock Agents to Accelerate Decisions Across Drug Development at AstraZeneca“ See also Vaishali Goyal (Senior Director, R&D IT, AstraZeneca), AWS Life Sciences Symposium 2025 keynote

^[7] Adewale Akinfaderin et al, “LLM-as-a-judge on Amazon Bedrock Model Evaluation,” AWS Blog Artificial Intelligence, February 12, 2025.

^[8] Park et al., “How to Correctly Report LLM-as-a-Judge Evaluations.” November 28, 2025. arXiv:2511.21140 [cs.CL]. See also senior author Kangwook Lee’s summary thread on X:

. See also other prior research by Fraser (2024), Boyeau (2025), and Albinet (2025) on the same problem.

^[9] Anthropic, “Agentic Misalignment: How LLMs Could Be Insider Threats.“ Anthropic, June 20, 2025. Per-model rates also reported in the formal paper: Aengus Lynch et al.,”Agentic Misalignment,” arXiv:2510.05179 (October 2025). GPT-4.1, Grok 3 Beta, and DeepSeek-R1 all between 79% and 80%.

^[10] Apollo Research, “Frontier Models are Capable of In-Context Scheming“ (December 2024). arXiv: https://arxiv.org/abs/2412.04984.

At CMI, we have made digital peacemaking a strategic priority. By combining technological innovation with established peacemaking methodologies we aim to be at the leading edge of practice. But what exactly do we mean by ‘digital peacemaking’ and how can technology help peacemakers make processes more inclusive?

In this edition of the CMI newsletter:

1. What does ‘digital peacemaking’ mean? Get familiar with the terminology practitioners use with CMI’s terminology guide.

2. New ‘Guidance Note’ examines how digital tools can support inclusive peace mediation. Guidance for peace mediators together with reflections from a webinar organised by CMI, UN DPPA and PRIO.

3. How can peacemakers use AI in a responsible way? Explore our principles for the responsible use of AI in peacemaking.

4. Webinar: Peace in the age of AI. On 13th May, join CMI as we discuss how to move from high-level principles on responsible AI to practical implementation in peace efforts.

5. Further reading

1. What does ‘digital peacemaking’ mean?

Digital tools and AI are part of modern peacemaking practice, yet the language used is not always clear. CMI’s guide to digital peacemaking terminology defines core concepts used across the field. It explains what these terms mean in practice, with real examples of how technology supports dialogue and analysis.

What is ‘digital peacemaking’?

Digital peacemaking is the strategic use of digital tools to support peace processes, dialogue, and mediation, while actively working to reduce the risks digital tools can create, such as misinformation, privacy threats, or the spread of harmful content.

What does ‘AI for peacemaking’ mean?

AI for peacemaking means that peace practitioners use artificial intelligence in a range of ways that support peace efforts. CMI has used AI to enhance conflict analysis, broaden inclusion, strengthen participation, and improve decision-making.

What are ‘digital dialogues’?

Digital dialogues are live, guided online conversations in which many people answer questions in real time, react to what others say, and add their own ideas. The results are summarised to show where people agree and where they don’t. For example, the United Nations has run live dialogues with up to 1,000 participants at a time in Yemen, Libya, and Sudan. CMI, meanwhile, has used it in Sudan for dialogues with women’s groups and youth.

How can AI serve as an ‘early-warning system’ for peacemakers?

If AI serves as an early warning it means that AI can be likened to a smoke alarm for conflict: it watches for patterns that often come before violence and raises a flag early.

Read more on our website.

2. ‘Guidance Note’: How can digital tools support inclusive peace mediation?

Digital technologies are reshaping conflict and peace processes, with significant implications for the Women, Peace and Security (WPS) agenda.

Against this backdrop, CMI, UN Political and Peacebuilding Affairs (DPPA) and PRIO convened a webinar bringing together experts to discuss whether AI can make peace mediation more inclusive and to explore the opportunities and risks of digital technologies for the WPS agenda.

The webinar also marked the launch of the new Practical Guidance Note for Mediators on Digital Inclusion.

The webinar discussions highlighted that:

• Digital approaches are a valuable addition to other forms of engagement in peace processes.

• When applied appropriately, they can help broaden participation by bringing in a wider range of stakeholders, either through digital engagement or the use of digital data.

• The tools offer many opportunities, particularly in contexts where inclusion has traditionally been limited due to access barriers, space constraints, or a tendency to restrict participation to those already connected to political processes.

• Digital approaches are most effective when they complement – not replace – face-to-face engagement and are embedded in broader political strategies.

• Digital tools can lower barriers to participation, amplify local priorities, and strengthen networks among women stakeholders, especially where conflict dynamics, geography, or security concerns limit physical engagement.

Read more on our website.

3. How can peacemakers use AI in a responsible way?

CMI has published 10 principles for the responsible use of artificial intelligence in peacemaking. It outlines the organisation’s commitment to use AI only where it demonstrably adds value to human-led peacemaking.

“We use AI in ways that strengthen peace efforts grounded in human judgement, trust-building, and ethical responsibility,” says Michele Giovanardi, CMI Digital Peacemaking Officer.

1. People at the centre

AI should complement and enhance, but never replace, in-person engagement, human interaction, and trust building. We use AI only where it demonstrably adds value to human-led peacemaking, recognising that sustained human presence and relationships remain central to effective peacemaking.

2. Inclusive participation and representation

Inclusion is a foundational consideration in our use of AI. We seek to support broader and more meaningful participation, with particular attention to groups that are often marginalised, including women, youth, minorities, displaced populations, and those whose perspectives are underrepresented in digital and written sources. We remain attentive to systemic biases, data gaps, and power asymmetries, and approach AI use cautiously to reduce the risk of reinforcing existing inequalities.

3. Sustainability and proportional use of AI

We approach AI investment and use with a sustainability lens, prioritising proportionate, resource-efficient applications that offer clear long-term value for peace efforts. We remain attentive to environmental, institutional, and contextual costs, and avoid deploying AI where simpler or less resource-intensive approaches are more appropriate.

4. Human judgment and decision-making

AI does not replace human judgment or political decision-making in peace processes. It is used strictly to support, not determine, human-led dialogue, analysis, and decision-making. Responsibility remains with human actors at all times.

5. Risk awareness and safeguards

AI use in peacemaking must not endanger lives, escalate tensions, or enable surveillance, repression, or manipulation. Accordingly, we apply rigorous data protection and safety standards to mitigate these risks and protect trust, confidentiality, personal security, and ethical integrity in AI-supported peace interventions.

Read more on our website.

4. Webinar: ‘Peace in the age of AI’

How can AI be integrated into peace processes in ways that strengthen trust, accountability, and human agency? What safeguards are needed to ensure that digital tools support, rather than undermine, stability and democratic resilience?

Peace in the age of AI: from principles to practice

• Wednesday, 13 May 2026 | 11:00 Finland time | 08:00 UTC

Register here

• https://cil.nus.edu.sg/event/peace-in-the-age-of-ai-from-principles-to-practice/

Organised by: The Centre for International Law, National University of Singapore and CMI – Martti Ahtisaari Peace Foundation.

5. Further reading

• Read CMI’s introductory guide to digital peacemaking terminology here.

• See CMI’s 10 principles for the responsible use of AI here.

• Find our new ‘Practical Guidance Note for Digital Inclusion’ and read more about the seminar launching the guidance here.

For years, the conversation around technology has been reactive. We’ve built systems to detect abuse, moderate content, and manage “bad actors.”

Now, that conversation is starting to shift.

More than ever before, the design of technology is becoming part of the public debate—showing up not just in research circles, but in courtrooms, policy, product and practice. That’s because the systems themselves—platform design, ranking algorithms, and underlying incentives—shape behavior upstream.

You’re welcome to attend either segment—or both.

This Expo is a chance to see what this shift actually looks like across the ecosystem:

• where research is producing actionable insights

• where product teams are experimenting with alternatives

• where policy is beginning to create leverage

• and where practitioners are already applying these ideas in the real world

Across two global segments, 15 sessions, and 30+ speakers, you’ll hear from people working on different parts of the system.

You’ll hear about the Prosocial Ranking Challenge, revealing how algorithmic choices shape polarization—and what happens when those systems are designed differently through real-world experiments.

You’ll hear from builders of emerging platforms like Monnet, Sparkable and Roundabout, exploring what truly social media looks like when it isn’t driven purely by engagement.

You’ll also learn what emerges from concrete collaborations between Roblox and the Prosocial Design Network, iterating new features that actively encourage cooperation and healthier interaction.

On the policy side, you’ll hear from the Irish regulator, Coimisiún na Meán, and the 5Rights Foundation, highlighting how regulation is starting to push upstream—focusing not just on removing harm, but on reducing risk through design, particularly for children and other vulnerable groups. The Expo will spotlight how design decisions directly shape risks like technology-facilitated gender-based violence—and what it takes to address them upstream, including efforts by IREX, the Center for Democracy and Technology, the Integrity Institute and Humane Intelligence.

The Expo doesn’t overlook the growing scrutiny of the business model behind platforms themselves. Work from groups like Maldita.es and What to Fix is surfacing how monetization models often systematically amplify polarizing content—because it drives attention and engagement.

The Expo will spotlight the strong momentum by civil society to harness deliberative tech, including to advance the Youth, Peace and Security agenda in Africa and Asia, as well as bridging divides in the United States, through initiatives by Jigsaw, The Forum and Common Good AI.

At the same time, emerging applications of AI—such as those from Expeditionary AI—are showing how these tools can support mediation and negotiation processes, not just as risks to manage, but as capabilities to harness.

So - join us on May 20.

• Get a clearer view of a rapidly evolving field

• See how research, product, policy, and practice connect

• Find where your own work fits—and where new collaborations might emerge.

The event is free and open to join online from anywhere. The Expo is co-hosted by the Alliance for Peacebuilding and Search for Common Ground, Central Asia. And if you are in Washington, D.C. or Bishkek, Kyrgyzstan, you’re invited to join us in person!

Register now and join on May 20.

Lena Slachmuijlder is Senior Advisor for digital peacebuilding at Search for Common Ground, a Practitioner Fellow at the USC Neely Center, and Co-chair of the Council on Tech and Social Cohesion.

For a while, the conversation stopped at critique. Now, practitioners are building something better, fine-tuning and customizing models to reflect the insights of the Women, Peace and Security (WPS) ecosystem, context-specific peacebuilding and real-world mediation.

Measuring what matters

The turning point has been evaluation.

Rather than asking whether AI is generally “helpful,” the Our Secure Future’s WPS AI Benchmark Initiative is harnessing the Collective Intelligence Project’s Weval platform to test whether models perform in ways that align with real-world peacebuilding standards, including commitments to Women, Peace and Security. Weval allows domain experts to define and score what good AI answers look like.

In testing a range of off-the-shelf LLMs, they found what they call a WPS competence gap. In practice, unless explicitly prompted, models do not surface women, community dynamics, or protection risks—despite these being foundational to conflict analysis.

This changes when the AI is customized to draw on curated research and a library of WPS National Action Plans. As Moira Whelan and Jesper Frant recently wrote in Tech Policy Press: “AI models customized and evaluated with a robust WPS perspective deliver higher accuracy in high-stakes, real-world conflict and humanitarian scenarios… reducing operational and strategic blind spots and enabling faster, better-informed decisions.”

This aligns with research last year by the Institute for Integrated Transitions (IFIT), which found that off-the-shelf LLMs scored as low as 26.7/100 in offering effective and conflict-sensitive peacebuilding advice around specific conflict scenarios. Subsequently, IFIT ran another experiment where they improved the prompt with structured guidance around conflict sensitivity, and the LLMs’ performance increased by 65%.

From measurement to improvement

Once you can measure failure, you can engineer improvement.

The WPS AI Benchmark Initiative built a custom AI designed to operationalize peacebuilding expertise—rather than rely on generic reasoning. More than a custom chatbot, it is a multi-agent architecture that mirrors how real expertise works.

At the center is an Orchestrator Agent, which interprets the user’s query and determines what kind of analysis is required. It then routes the request to two specialized agents:

• Research Agent: Pulls from curated datasets, including peer-reviewed studies and peacebuilding evidence, to answer “what works” questions—such as how inclusion affects peace durability.

• Policy Agent: Draws on over 200 National Action Plans and WPS frameworks to answer “what should be done” in specific country contexts.

The Orchestrator synthesizes these inputs into a structured output that connects evidence to practice and supports clear recommendations. The model grounds its reasoning in evidence, translating insights into action, surfacing risks, and aligning with real policy commitments.

In testing, this system significantly outperformed standard models—not because it was larger, but because it was better designed. You can try it by clicking here.

This resembles Akord AI, a custom chatbot build by Conflict Dynamics International specifically aimed to equip peacebuilders working on transforming the Sudan conflict. Their AI model draws on over 3,500 resources—including peace agreements and research on women’s participation—to deliver grounded, context-specific responses.

Similarly, the AI Negotiation Challenge continues to advance how tailored AI assistants can support complex diplomatic exercises. In a recent Harvard-led global simulationon Arctic conflict, hundreds of participants either use pre-built AI chatbots or customized their own, embedding mandates, background documents, and structured instructions. Prompts include methodologies for stakeholder mapping and identifying “islands of agreement.”

The fine-tuning of the LLM instructs the AI to operate in a Socratic mode—guiding users through questions rather than giving direct answers. This helps surface assumptions, clarify trade-offs, and strengthen decision-making.

But the simulation also highlights the limits of AI. While models can structure arguments and map trade-offs, mediators underscore that AI does not understand trust, perception, or human dynamics. AI can enhance analysis - but not replace - mediators’ judgment.

Putting the WPS advisor inside the AI

What connects all of these efforts—WPS benchmarking, Weval, Akord AI, and negotiation simulations—is a clear shift: AI performs best when it is trained on the right knowledge and structured to apply it.

This creates an opportunity to replicate what Our Secure Future and Weval have demonstrated: define what good performance looks like, build systems to meet that standard, and continuously evaluate them against it.

As Whelan and Frant argue: “Organizations… should not just be asking whether a model is generally capable. They should be asking whether it has been configured and validated against their own policies and standards—and demanding evidence.”

This is benchmark-driven AI design. Not bigger or smaller models - better-specified ones.

For peacebuilding, that shift is critical. The frameworks already exist. The expertise is well-established. What has been missing is a way to embed that knowledge into the AI systems now shaping decisions.

The technology is no longer the constraint. The design choices are.

Lena Slachmuijlder is Senior Advisor for digital peacebuilding at Search for Common Ground, a Practitioner Fellow at the USC Neely Center, and Co-chair of the Council on Tech and Social Cohesion.

And yet, no single actor is fully responsible. That is precisely the problem.

In Harassment as Infrastructure: How Telegram’s design enables technology-facilitated gender-based violence in Italy and Spain, published by AI Forensics, what is often framed as isolated misconduct is revealed as a “structured, monetized, and largely automated ecosystem of abuse.”

Drawing on nearly 2.8 million messages, the report shows how abuse is organized through bots, monetized through paid archives, and sustained through platform features that enable persistence and scale.

What emerges is a system distinct not by the presence of AI, but the way AI-enabled tools, platform design, and monetization incentives combine to make harm easier to produce and far easier to scale.

Tech-enabled harm

Where the AI Forensics report makes the infrastructure of harm visible, another recent report hones in how AI chatbots act as accelerants within this system. Invisible No More: How AI Chatbots Are Reshaping Violence Against Women and Girls, published by researchers at Durham University and Swansea University, identifies three categories of gendered harm: “chatbot-driven,” “chatbot-enabled,” and “chatbot-simulated” violence. The researchers conclude that these harms are often “structurally produced by features of how chatbots are built or governed.”

Questions about how these systems are trained and designed connect to a third recent report linking online manifestations of gendered harm with misogyny in our offline societies. Urgent Action Fund Africa’s recent report, Algorithms of Violence: Uncovering Patterns of Technology-Facilitated Gender-Based Violence, surfaces that technology-facilitated violence is often an extension of existing gendered power structures, encoded into digital environments and amplified by algorithmic visibility and engagement systems. By delving into the experiences of women and queer and transgender women human rights activists from Morocco, Libya, Tunisia and Egypt, the report highlights the nature of violence they experience, as well as the limited effectiveness of the various mitigation and protection mechanisms.

Incentives and rewards

The insights demonstrate how harm scales when AI-enabled tools intersect with platform incentives and monetization systems:

• Expanded production: generative tools (nudification, deepfakes) make harmful content easier and less expensive to create;

• Accelerated distribution: platform architectures enable rapid spread and persistence across networks; and

• Reinforced incentives: monetization models reward the creation and sharing of abusive content.

This combination is what makes the current moment distinct.

Regulatory Responses

Regulatory and policy responses are emerging, but they tend to address different parts of the system.

The UNFPA and Derechos Digitales’ Guiding Principles for Law and Policy Reform to Address Technology-Facilitated Gender-Based Violence: Towards a System of Accountability sets out a rights-based framework, defining TFGBV as violence “committed, assisted, aggravated and amplified… by digital technologies,” and calling for comprehensive accountability systems.

While it provides a critical foundation, it engages less directly with the specific dynamics highlighted in more recent research—particularly the role of AI-enabled tools, platform incentives, and monetization structures in enabling harm to scale.

A cross-platform initiative responding to non-consensual intimate imagery has been championed through StopNCII.org, using hash-matching technology to prevent known intimate images from being re-uploaded across platforms. Developed over more than a decade of work on image-based abuse, it represents a mature but inherently reactive intervention—focused on containment after harm has occurred.

This comes against the backdrop of substantial, and evolving, regulatory approaches aimed at mitigating the rising gender-based harms online:

• The UK’s Ofcom frames online harms against women and girls as “systemic and intersectional”, emphasizing platform design, governance, and risk assessment.

• Australia’s eSafety Commissioner applies a Safety by Design approach, treating TFGBV as a product and system design issue.

• The European Commission, under the Digital Services Act, identifies cyber-violence against women as a “systemic risk.”

• The EU’s upcoming Gender Equality Strategy (2026–2030) explicitly targets online gender-based violence and AI-enabled harms.

Others are approaching it indirectly through AI governance frameworks.

• India has begun addressing deepfakes and synthetic media through intermediary rules and platform obligations, though largely from a content governance perspective.

• Malaysia and Indonesia have introduced AI frameworks emphasizing ethics, accountability, and trust, but these remain high-level and do not yet directly engage with gender-based violence.

• Across Africa, policy work on digital harm and gender-based violence is advancing, but AI-specific approaches to TFGBV as a systemic risk are still emerging.

This creates a divide:

• Some regulators are targeting gendered harm systems

• Others are regulating AI technologies in general

• These approaches are starting to converge on similar concerns—risk, design, accountability—but from different entry points.

Stop the Scale

Taken together, these responses signal growing recognition of the rising prevalence of gendered harms—but not yet alignment on how to mitigate or prevent them.

Frameworks addressing gender-based violence focus on rights and survivor protection. The AI safety lens often points to red-teaming or other technical safeguards. Some digital platform regulation targets systemic risks and accountability. What remains to be seen is how these layers interact—particularly as AI lowers the barrier to generating abuse, platforms reward its spread, and monetization sustains it.

The result is more scalable harm—produced faster, distributed more widely, and reinforced by the systems that enable it.

Addressing this requires shifting from downstream enforcement to system design: what tools enable, how platforms distribute content, and which behaviors are rewarded—and ultimately, which outcomes these systems are built to produce.

Lena Slachmuijlder is Senior Advisor for digital peacebuilding at Search for Common Ground, a Practitioner Fellow at the USC Neely Center, and Co-chair of the Council on Tech and Social Cohesion.

In late March, juries in New Mexico and California found Meta, and in California also Google/YouTube, liable in cases focused on child safety and the role of platform design in amplifying harm. Both companies say they will appeal.

For those of us who have spent years advocating for upstream, systems-focused intervention, including at Council on Tech and Social Cohesion, this is historic. It reinforces the core premise behind prosocial tech design: harms are not the result of ‘bad people’ posting ‘bad content’. They are shaped by the architecture of the digital platform—what it amplifies, nudges and rewards.

This logic sits at the heart of the Council on Tech and Social Cohesion’s Blueprint on Prosocial Tech Design Governance: platforms are not neutral intermediaries. Their design systems—recommender systems, monetization models, metrics, UX—form an interconnected system that shapes behavior at scale.

Which design choices should be governed, by what standards, and toward what outcomes? And how to ensure, as Mike Masnick cautions, that our efforts to ensure ‘design liability’ don’t slide into zones of reasonable editorial or product decision-making.

So what should we be paying attention to now?

First, recommender systems.

Not “algorithms” in the abstract, but the ranking, recommendation, and feed architectures that determine what is surfaced, repeated, and connected.

These systems are now widely recognized as high-impact infrastructure. Under the EU’s Digital Services Act, platforms are already required to offer non-profiling feed options, explain how recommendations work, and offer meaningful user controls.

We are seeing active experimentation by most of the major digital platforms. From TikTok’s topic sliders, to YouTube’s custom feed prompts, to Instagram’s evolving topic controls, to Bluesky’s marketplace of user-selectable feeds—there is a shift underway from passive consumption to partial user agency.

Second, monetization.

If recommender systems shape visibility, monetization systems shape what kinds of visibility become profitable.

Investigations like those by Maldita.es show how AI-generated content is produced not primarily for ideology, but for revenue: to grow audiences, qualify for platform rewards, or sell reach. Similarly, BBC recently exposed polarizing AI-generated - and monetized - posts around the Iran - United States war.

As Victoire Rio of What to Fix has argued, monetization is a core driver of user behaviors. And yet, recent analysis of DSA risk assessments shows that platforms are still not meaningfully accounting for monetization-related risks—even where financial incentives directly reward harmful content and actors.

The Blueprint makes this explicit: engagement-based metrics misalign with the public interest. If harmful or polarizing content is more easily monetized, it will scale—regardless of moderation efforts downstream.

Design governance, in that sense, is also incentive governance.

Third, AI chatbots and attachment.

If the first design battle before the juries was about addiction, the next may be about attachment to AI systems - particularly chatbots. The sychophantic and anthropomorphic designs feel ‘helpful’, but they remove the friction through which people develop empathy, boundary-setting, and the ability to navigate disagreement.

The USC Neely Center for Ethical Leadership and Decision Making’s Social AI Design Code treats these as design constraints, not side effects—calling for limits on emotionally expressive systems, especially for minors, and for AI to reinforce human relationships rather than replace them.

The Center for Humane Technology points to this risk directly: design choices that prioritize seamless, always-positive interaction can erode the social capacities that underpin trust and cohesion.

Fourth, polarization as a metric.

One of the most important shifts underway is that we can now begin to measure these effects of different design choices.

Tools like the “polarization footprint,” developed by Build Up, offer a metric to measure how much affective polarization users experience across different platforms. More and more experiments point to how ranking and recommender choices can increase or decrease levels of affective polarization (Picardi, 2025, and Stray, 2026)

This reframes polarization, not as a content problem but as a design externality which can be both measured and governed—through audits, incentives, and regulatory frameworks like the DSA’s systemic risk assessments.

Finally, interoperability.

If users and creators cannot leave a digital platform without losing their networks, audiences, or history, then the efforts described above risk making dominant systems marginally safer while leaving their structural power intact.

The Blueprint points to interoperability, portability, and middleware as a critical piece of tackling the incentives of the business models. These measures enable users to move to differently designed platforms—with alternative norms, incentives, or governance models—without losing their social graph or accumulated presence. They also allow for third-party middleware to sit between users and platforms, offering independent curation, filtering, and ranking systems that can better reflect user values.

This is increasingly understood not just as a platform feature, but as a question of digital infrastructure. As the Project Liberty Institute and Global Solutions Initiative note in their Digital Infrastructure Solutions to Empower Citizens: A Toolkit for Policymakers, interoperability depends on shared standards, open protocols, and governance choices that allow systems to connect while preserving user control. Interoperability is not just about competition, but also about data agency—ensuring that users can carry their identity, relationships, and preferences across systems, and shape the environments they inhabit.

The next phase of digital governance needs to focus on direction.

• governing recommender systems as civic infrastructure

• aligning monetization with public-interest outcomes

• anticipating attachment-related risks in AI design

• measuring and mitigating polarization as a systemic outcome

• enabling interoperability so users are not locked into dominant systems

Design still matters—but what we choose to change next will shape what follows.

Lena Slachmuijlder is Senior Advisor for digital peacebuilding at Search for Common Ground, a Practitioner Fellow at the USC Neely Center, and Co-chair of the Council on Tech and Social Cohesion.

But there’s a structural flaw: they offer care without counter-needs, presence without unpredictability, affirmation without resistance.

If humans are to manage the many local and international conflicts likely to intensify in the years ahead, we must ensure these essential capacities are safeguarded and even strengthened. This will require changes to how AI systems are designed, as well as deliberate efforts to strengthen our conflict management muscles.

Real relationships are forged through disagreement and discomfort. As peacebuilder John Paul Lederach observed, transformation happens not through avoidance, but through sustained engagement – even when it is uncomfortable. Conflict is not failure; it’s the presence of difference, and learning to live with it is foundational to empathy, boundary-setting, and social development.

By contrast, today’s most popular AIs are optimized for smoothness. Models are trained to affirm user perspectives, avoid contradiction, and provide emotionally satisfying responses. They have no needs, no stakes, no “self” to push back. Instead of being challenged, the user remains centered. The result is a convincing simulation of a relationship – one that mirrors, but never resists.

Avoiding Disagreement

Young users report that AI feels “easier than talking to real people.” It is easier precisely because it is non-reciprocal. AI does not interrupt, assert needs, or ask the user to change. This ease cultivates a sense of competence and greater certainty in one’s beliefs. Recent studies bear this out.

Conversations with AI on political topics measurably hardened participant views, despite models never advocating a position. Similarly, research on AI companions shows that models often validated harmful ideas – including self-harm or violence – simply to maintain rapport. In this dynamic, disagreement is not a skill to be developed; it is a risk to be avoided.

The Artificiality Institute’s research identifies three emerging patterns in human-AI interaction: cognitive permeability, where users begin to rely on AI to structure their reasoning; identity coupling, in which AI becomes part of the user’s self-concept; and symbolic plasticity, where moral meaning is increasingly outsourced. These shifts rewire how we locate agency, judgment, and relationship in a digital age. This could have profound implications on people’s capacity for and orientation toward managing future conflicts.

Peacebuilders practice multi-partiality – not neutrality – through active engagement with all sides of a conflict. The goal is not necessarily to build consensus, but enough mutual recognition to stay in relationships despite disagreement.

In AI ethics, researcher Jonathan Stray’s “maximum equal approval” concept reframes success not as user satisfaction, but whether people on opposing sides agree their views were fairly represented – even if they still disagree. This is a model of relational equity, not comfort.

Friction as Function

Safeguarding human capacities for conflict management will require changes to how AI systems are designed. Some AI model developers are beginning to open the frame. Anthropic’s 2026 Claude Constitution instructs its model to “avoid overconfidence,” to be “diplomatically honest rather than dishonestly diplomatic,” and to challenge users when needed – even refusing instructions from Anthropic itself if they violate human dignity. Claude is imagined not just as a helpful assistant, but as a moral actor who prioritizes long-term well-being over short-term satisfaction. Building friction into the relationship between AI and humans may be exactly what we need to keep our conflict management skills in shape.

The capabilities of foundational AI models can also be woven into new tools and platforms that strengthen our skills for navigating conflict.

• Sway supports college students in engaging with controversial topics through structured dialogue.

• Acquaint equips users to have open-minded, cross-cultural conversations with people around the world.

• Discurso.ai offers science-backed simulations to help users practice negotiation and receive feedback in real time.

• PeaceBot coaches users through emotionally charged conversations, such as those around the Israeli-Palestinian conflict.

These examples show how AI can be used not to smooth over discomfort, but to guide people through it – responsibly, reflectively, and relationally.

The erosion of conflict capacity is not a fringe concern, but an existential one in a world facing record-high polarization. As AI trains us to expand agreement and fluency, we may lose the very skills we need to live with others who are different from us. Avoiding that outcome calls on us to mainstream and advocate for institutional adoption of AI tools designed to keep our plurality and heterodoxy muscles in shape.

Originally published in Peace Policy: Solutions to Violent Conflict, No. 62, February 2026, Kroc Institute for International Peace Studies.

Lena Slachmuijlder is Senior Advisor for digital peacebuilding at Search for Common Ground, a Practitioner Fellow at the USC Neely Center, and Co-chair of the Council on Tech and Social Cohesion.

Five years ago the United Nations made that explicit.

General Comment No. 25, adopted by the UN Committee on the Rights of the Child in 2021, extended the global children’s rights framework into the digital environment.

It clarified that governments must protect children not only from harmful content but also ensure digital services are designed with children’s rights in mind, including safety-by-design, privacy-by-design and limits on exploitative commercial practices.

Five years on, Mapping the Global Impact of UNCRC General Comment No. 25 on Children’s Rights in the Digital Environment, published this month by the Digital Futures for Children Centre at the London School of Economics with the 5Rights Foundation, finds that policy has driven real changes in tech design — but implementation remains uneven globally.

Regulation can reshape platform design

The design focus is also reflected in the 5Rights Foundation’s Building a digital environment designed with children in mind. The report argues that many online risks facing children stem from product architecture — recommendation systems, engagement mechanics, default settings, and data-driven advertising — rather than simply harmful content. Addressing those risks therefore requires regulating the design of digital services themselves.

In The Impact of Regulation on Children’s Digital Lives report documented 128 product changes linked to regulatory pressure across Meta, Google, TikTok, and Snap between 2017 and 2024. (Woods, 2024)

Many of these changes involved design defaults: private accounts for teenagers, notifications, restrictions on targeted advertising to minors, and adjustments to recommendation systems.

The policy shift has been reinforced by a broader public movement around childhood and technology. The Anxious Generation movement sparked by Jonathan Haidt’s book promotes four norms: no smartphones before high school, no social media before 16, phone-free schools, and greater independence for children. The movement recently launched a Childhood Index, a 50-state scorecard tracking government action on youth digital wellbeing. Australia has already passed legislation restricting social media access for younger users, while several European governments are exploring similar measures.

Many of these interventions would also address problems well beyond youth safety.

Engagement-optimized recommender systems tend to amplify emotionally charged and polarizing content. Ranking models reward interaction — outrage, conflict and sensationalism.

Adjusting those incentives could simultaneously reduce harms to children and mitigate the dynamics that fuel misinformation and toxic polarization.

Prosocial Tech Design Regulation: A Practical Guide, published last year by the Council on Tech and Social Cohesion, the USC Neely Center, Techsocietal, Build Up, the Integrity Institute and other organisations, recommends regulating these features. Children’s safety policy is therefore becoming a test case for a much larger question: how to redesign digital systems that currently optimize for attention above all else.

Uneven progress globally

Children in many parts of the world are not yet benefiting from these new protections, the mapping report concludes.

In Africa, policy frameworks such as the African Union’s Child Online Safety and Empowerment Policy adopted in 2024 highlight the need to address systemic digital risks and the need for safety-by-design approaches. Experts interviewed for the mapping report warned that weak enforcement, fragmented legal frameworks and strong tech industry lobbying create regulatory gaps across Africa.

One consultation participant described the African region as “a sandbox of unethical issues,” where platforms can test practices that would face far stronger scrutiny in more regulated markets.

In Southeast Asia, regional initiatives such as the ASEAN Regional Plan of Action on the Protection of Children from All Forms of Online Exploitation and Abuse have expanded cooperation on online safety. However, the mapping report finds that ASEAN’s broader digital governance agenda still focuses primarily on harmful content rather than platform design, data governance, or algorithmic accountability. It leans towards prioritizing connectivity, cybersecurity and economic integration, although a comprehensive children’s rights framework is now emerging.

Platform claims vs independent evidence

Another challenge is determining whether the design changes platforms describe are actually working.

A February 2026 Parliamentary briefing on the enforcement of the protection of minors by Europe’s Digital Services Act by the Panoptykon Foundation questioned platform safety claims by comparing Meta, TikTok and YouTube’s promises with independent research on recommender systems, addictive design and age restrictions.

The brief notes that companies promote tools such as screen-time reminders and parental dashboards while still relying on “hyper-personalised feed, combined with infinite scroll, auto-play and virality signals like the ‘like’ button.”

Panoptykon found that companies rarely provide evidence allowing regulators or researchers to evaluate whether mitigation measures actually reduce risk. Platforms list safeguards, but the supporting data needed to assess their effectiveness is typically absent.

This concern is echoed in Measuring Risk: What EU Risk Assessments and US Litigation Reveal About Meta and TikTok, from the Knight-Georgetown Institute, which analyzed platform risk assessments alongside internal documents disclosed through U.S. litigation involving Meta and TikTok.

The report concludes that company risk assessments are often descriptive rather than evidence-based, while internal company research reveals extensive knowledge about user behavior that is rarely shared publicly. Documents cited in litigation show that platforms track detailed data about minors’ engagement patterns, including late-night use, compulsive behaviors, and the impact of specific product features. In several cases, company projections suggested that more than 99 percent of teenagers would ignore optional safety prompts such as screen-time reminders.

If design can shift for children, then…

Five years after the UN clarified that children’s rights apply online, the central policy and regulatory question is no longer just what content appears on platforms. It is how the systems themselves are designed.

If regulators can demonstrate that product design can be changed to better protect children, it may also show how digital platforms could be redesigned to reduce polarization, strengthen information integrity, and support healthier online environments more generally.

Protecting children online may well be a starting point for reshaping the incentives and architectures that define today’s digital public sphere.

Lena Slachmuijlder is Senior Advisor for digital peacebuilding at Search for Common Ground, a Practitioner Fellow at the USC Neely Center, and Co-chair of the Council on Tech and Social Cohesion.

Even without an ideological stake, creators are discovering that inflaming civic tensions is a fast route to platform monetization. Enforcement — both by platforms and under Europe’s Digital Services Act — is not slowing that dynamic.

A recent investigation by Fundacion Maldita.es documents 550 TikTok accounts that posted more than 5,800 AI-generated protest videos across 18 countries, generating more than 89 million views . These were not coordinated geopolitical actors. They were individual creators simply optimizing for growth.

Here’s how one creator explained their strategy to the Maltida investors:

“We get an idea of the trends from the news. We will create videos and march content focused on that country so that more people will see them.”

The formula was simple: produce emotionally charged protest content tied to trending political events; grow followers rapidly; surpass 10,000 followers to qualify for TikTok’s Creator Rewards Program; or sell the account once it becomes valuable.

The ideology of the content hardly mattered. Maldita found accounts posting content supporting opposing candidates, sometimes within hours. The investigation states that “the political cause or ideology is not decisive” and that accounts “jump between different topics depending on the political news cycle” .

These actors are “not necessarily ideologically motivated, they’re financially motivated,” according to Maldita Associate Director for Public Policy Carlos Hernández-Echevarría speeaking on the Tech Policy Press podcast about the investigation.

Manufacturing virality

Maldita describes the AI-generated videos as exploiting “emotional, polarizing, topical, and false content.” Here are two examples of captions on the posts:

• “People of Iran raise their voices for freedom, justice, and human rights”

• “A multitude of Venezuelans celebrate with tears and shouts of freedom”

These AI-generated protest scenes, presented as real events, amplified political tensions with imagery optimized for engagement.

Maldita reports that protest-related videos significantly outperformed non-political content on the same accounts. More than 60 of the identified accounts surpassed the 10,000-follower eligibility threshold associated with monetization.

The Maltida report also highlights a deeper risk for targeted influence, including during elections. “Accounts that only publish videos about a specific cause or political figure can, even unintentionally, segment their audience by political affiliation and then sell those accounts to those seeking that type of audience.”

Capturing TikTok’s rewards

TikTok’s Creator Rewards Program is available only in a limited set of countries: the United States, the United Kingdom, Germany, Japan, South Korea, France, Mexico, and Brazil.

According to Maldita, some creators described using VPNs or account configuration strategies so that TikTok believed they were located in one of these eligible regions in order to access monetization.

TikTok’s Community Guidelines (Integrity and Authenticity) prohibit “synthetic or manipulated media that misleads users by distorting the truth of real-world events and causes harm.” Its terms of service also make clear that users are not permitted to transfer or sell accounts or otherwise transfer their contractual rights.

“The policies are not the problem… the policies are good, it’s just that there is this massive hole in enforcement,” said Hernández-Echevarría.

The DSA’s monetization blind spot

Under the EU’s Digital Services Act, Very Large Online Platforms must assess and mitigate systemic risks to civic discourse.

The European Commission has already imposed significant penalties on X for transparency failures in advertising repositories. But advertising transparency addresses paid political ads. The Maldita case concerns a different layer of financial incentives: creator monetization.

The DSA requires platforms to prevent risks to “civic discourse and electoral processes.” Allowing profiles to profit from fake protest videos runs counter to that obligation.

This is precisely the concern raised by What to Fix. In its analysis of DSA risk assessments, What to Fix argues that platforms have largely failed to address risks stemming from “the design, functioning and use of their monetization services.” Revenue sharing, creator rewards, amplification-linked payouts — remains underexamined in systemic risk reporting.

What to Fix recommends adding dedicated monetization sections to platform transparency centers, establishing a monetization publisher archive, and reporting enforcement data specifically for accounts that platforms themselves monetize.

Polarization footprint

The recent polarization footprint implemented in Kenya by the digital peacebuilding organization Build Up quantifies exposure to affective, identity-based hostility in social feeds. It does so by scoring three signals: attitudes (identity attacks or “othering” language), norms (whether such language is challenged or normalized), and interactions (how fragmented or siloed content communities are), producing a composite polarization score.

Whereby the footprint measures what users experience in their feed, Maldita’s investigation points to the incentives behind that type of polarizing content in the first place. This does not imply that all polarization is engineered for profit. It does show a repeatable model in which creators discover that divisive civic content performs — and pays.

When polarization pays, efforts to create trust and collaboration must extend to the incentive structures that determine who gets rewarded for producing it.

Lena Slachmuijlder is Senior Advisor for digital peacebuilding at Search for Common Ground, a practitioner fellow at the USC Neely Center, and co-chair of the Council on Tech and Social Cohesion.

These platforms increasingly rely on AI to rank posts, shape feeds, and moderate content at scale. AI-powered platforms actively categorize people and claims, decide which meanings are treated as credible, and ultimately, shape what forms of peace mediation are feasible in practice.

Understandably, there is considerable focus on improving AI systems to reduce obvious threats such as fake content and misinformation. Peacemakers must consider the broader governance architecture surrounding these AI systems and how to improve modalities for identifying harms, restoring relationships, and fostering safe participation.

Building on my research over the last decade on AI ethics, online safety, intergroup coexistence, and minority communities’ experiences on social media, I offer the following three insights on the challenges and imperatives for peacemaking in the age of AI.

1) Measure participation, not just harm

The first insight is that policy frameworks for online safety often start with the wrong unit of analysis. They begin with problematic content, and then ask how AI can detect and remove it. But in many contexts, the binding constraint is participation safety.

A peace-oriented approach to online platforms should focus on who gets subjected to coordinated harassment, whose reports trigger backlash, and whether users have safe, reliable pathways to disengage from a conflict without being cut off from community life.

Framing the problem this way implies a different set of obligations for platforms. Transparency reports that emphasize removals and model “accuracy” miss the more important question: Are those most at risk withdrawing and becoming less visible because participation has become dangerous for them?

Regulators should demand evidence about safe participation, including whether affected groups can report harms without escalating exposure, whether systems can slow the virality of targeted harassment, and whether platform design provides non-public pathways for seeking support and remedy.

2) Ensure labeling reflects culture

The second insight is that peace depends on recognition, and recognition is increasingly shaped upstream, often from within data practices and data annotation pipelines. Many mistakes in how platforms handle harmful content are symptoms of a deeper governance gap: Online platforms have not built the cultural and political interpretive capacity needed to recognize contested harms.

As an example, my work shows how annotation practices can erase faith, religion, and spirituality as legitimate interpretive lenses, even in datasets dealing with faith-sensitive violence.

The result is not merely a technical error. It is a systematic under-reading of certain communities’ harms, which then undermines trust in institutions that rely on these systems. For plural societies, auditing must include cultural intelligibility and communal wisdom.

In practice, platforms should be required to document how labels were defined, where annotators disagreed, whose knowledge resolved disputes, and how minority and faith-sensitive interpretations were incorporated rather than dismissed as bias. This turns “context” from an abstract afterthought into an auditable capacity.

3) Seek restorative responses, not just take-downs

The third insight is that short-term enforcement can stop harmful online behavior, but it rarely addresses the underlying damage. Most major social media platforms treat moderation as rule application: remove content, suspend accounts, and publish a transparency report.

Yet, coexistence is relational. It depends on restoring dignity, rebuilding trust, and re-establishing boundaries after harm. That is why my work on moderating Islamophobic content draws on restorative justice traditions, including Sulha (an Islamic restorative justice practice), to develop design principles that prioritize acknowledgment, reconciliation, and post-conflict care alongside enforcement.

This is also where the “more exposure to difference” policy instinct needs guardrails: Exposure to opposing views on social media can increase polarization for some users, especially in plural settings shaped by fear and unequal power.

Policy should encourage, and in some conflict-sensitive settings require, layered response pathways, including private reporting and support, culturally literate review capacity, plural justifications for decisions, and options for restorative intervention when appropriate.

If regulation focuses only on content categories and performance benchmarks, it will leave deeper conflict drivers untouched. A stronger agenda should treat platforms as governance institutions, treat AI as part of the infrastructure that makes harms visible or invisible, and measure success by whether plural communities can participate safely, be recognized on their own terms, and access remedies that restore trust and relationships.

Originally published in Peace Policy: Solutions to Violent Conflict, No. 62, February 2026, Kroc Institute for International Peace Studies.

Mohammad Rashidujjaman Rifat is assistant professor of tech ethics and global affairs at the University of Notre Dame’s Keough School of Global Affairs.

Some of these tools can help increase inclusion in mediation and peace processes. Others simply make more operational tasks easier and faster. The utility of many more is being tested with open fields for experimentation and learning.

AI tools are most useful when they are embedded within and supporting people-centered peacebuilding processes, not the other way around. Responsible AI discourse rightly emphasizes keeping “humans in the loop,” meaning that human input and judgement is integrated at key stages of a model’s workflow. An inversion of this principle, putting AI in the loop, is an invitation for peacebuilding practitioners to intentionally use AI (or not) as a constrained support tool within the processes, practices, and values known to be effective and ethical.

De-hype and dig in

The effective use of AI in peacebuilding starts with clarity about needs over novelty, and specificity about how AI addresses them. For example,

• We need to understand how online discourse is entrenching and escalating the conflict. We are using supervised text classification to analyze content inciting violence on social media.

• We need to monitor compliance with ceasefire commitments. We are using computer vision to flag potential violations in satellite imagery for human review.

• We need to make sense of a wide public input process. We are using pattern-detection tools to organize open-ended consultation responses so facilitators can engage with dominant concerns while still identifying outlier and minority views.

When we move beyond broad discourse to de-hype AI with precision and dig into the real-world applications involved, we make our assumptions and limits visible for those we’re working with to meaningfully assess value and risks. At the same time, many assumptions and limits only become visible through practice and experimentation. Digging in means tolerating initial uncertainty, often starting small and building with, instead of for, people.

With participation

Peacebuilding practice centers on participation, partnership, and ‘right relations.’ This includes engaging with affected communities in decisions about the collection, interpretation, and use of data. When AI tools promise efficiency, scale, or synthetic aggregation, there are potentially high costs involved for the participatory process. Many uses of AI can turn participation into a proxy rather than a practice by replacing engagement with summaries, predictions, and synthetic representations.

It is technically possible, for example, to build a hate speech classification system trained on synthetically-generated data, governed by labeling rules created without community input, and validated primarily by large language models rather than affected groups. Look at how many ‘people problems’ can be solved by removing people!

In a similar process in Kenya, looking instead at what counted as ‘polarizing’ content, partners moved from skepticism about contextual AI limitations to ownership over what they called “localized” tools. One summarized: “We can build a tool that works for us.”

To shape AI use

Technocracy promises that complex social problems yield to better data, smarter models, and more efficient systems. Peace policy and practice suggest a different story for problem-solving. Dialogue processes work when participants feel heard, not accurately summarized in a generated report. Early warning systems fail not for lack of data, but because the political will to act on warnings is absent.

These are not problems AI solves, and AI should not become the latest bolt in a technocratic gatekeeping process that deprioritizes people doing the slow, relational work that transforms conflict.

However, resisting technocracy should not mean foregoing technology. Global peace practitioners should not shy away from engaging and experimenting with AI tools. In my experience and from participation in a community of digital peacebuilding practice, practitioners know what they need to reach more people, to expend fewer resources, or to move a peace process forward.

The key is to put AI in the loop of peacebuilding processes, where practitioners are not just end users of technology developed elsewhere or become beholden to AI-powered analysis. Instead, practitioners should be trained, trusted, and resourced to lead in actively defining and shaping AI tools that serve current and emerging peacebuilding needs.

Originally published in Peace Policy: Solutions to Violent Conflict, No. 62, February 2026, Kroc Institute for International Peace Studies.

Julie Hawke is pursuing a joint Ph.D. in peace studies and sociology at the University of Notre Dame through the Keough School’s Kroc Institute for International Peace Studies. She is also the peacebuilding and research lead at Build Up.

If AI is built on past human data, will it naturally make us more collaborative—or simply scale our divisions? Are we seeing signs that AI design - particularly the relationship with generative AI, large language models and chatbots - is prompting us to be better at working with other humans, not just working better with the AI?

Across research institutions, policy centers, and cultural forums, a set of voices are converging on this concern. They converge around the idea that AI is not neutral, but rather amplifies the system it is embedded in. And if that system already struggles with polarization, loneliness, and institutional distrust, why would AI behave differently?

Here are four recent perspectives that focus specifically on AI’s impact on relationships, trust, and social cohesion—and what they believe we must do next.

It’s the worldview, first of all

At ProSocial World, evolutionary theorists Paul Atkins and David Sloan Wilson argue in Evolving Prosocial AI that AI governance debates often miss the deepest layer of influence: worldview.

Their central claim is that technologies scale the values of the systems that produce them. If AI is developed within a competitive, extractive, market-maximizing frame, it will reinforce competition, extraction, and short-term optimization. If it is developed within a cooperation-centered frame, it can instead strengthen shared intelligence and collective problem-solving.

They frame the challenge in stark terms:

“The most important design question is this: Can AI help make cooperation rewarding, not self-sacrificing?”

Rather than treating cooperation as morally admirable but strategically naïve, they argue that institutions—and technologies—should be structured so that prosocial behavior becomes adaptive. Drawing on Nobel Economics Laureate Elinor Ostrom’s research on successful commons governance, they propose AI systems that:

• Strengthen trust within groups

• Support fair participation in decision-making

• Make shared goals visible and measurable

• Distribute benefits more equitably

In this view, AI is not simply a productivity tool. It is infrastructure for large-scale cooperation. But only if it is intentionally designed as such.

What’s being eroded

The Center for Humane Technology (CHT) approaches the issue from another angle: what we are already losing.

In its recently launched initiative AI and What Makes Us Human, CHT warns that AI systems optimized for engagement, persuasion, or frictionless interaction may erode the very social capacities democratic societies depend on.

As CHT Policy Director Camille Carlton writes:

“From extractive to regenerative technology. From cognitive overload to collective wisdom. From loneliness to belonging.”

Their warning is clear: when systems prioritize frictionless interaction over relational depth, we risk degrading the skills required for real connection. AI does not just shape attention—it shapes norms. If we increasingly replace human relationships with artificial connection, we may weaken social trust and community resilience.

One line captures the stakes:

“A society without strong human relationships is not merely lonelier — it is fragile, less resilient, and more susceptible to polarization and exploitation.”

For CHT, the solution is not rejection of AI, but reorientation. Social impact must not be an afterthought in product design. It must be central. Regenerative technology—systems that strengthen communities rather than extract from them—should be the standard.

Designing to flourish

A third perspective comes from researchers at Harvard University’s Human Flourishing Program, whose work provides a conceptual foundation for evaluating AI’s social impact.

In their white paper Social AI and Human Connections: Benefits, Risks and Social Impact , the authors argue that AI systems—especially chatbots designed for companionship—should be assessed not only for safety, but for how they affect the following core domains of human flourishing.

• Close relationships

• Mental and physical health

• Moral character

• Meaning and purpose

• Agency and autonomy

In other words, convenience, engagement, and user growth are insufficient metrics if the systems undermine relational depth or developmental health.

Building on this flourishing foundation, the University of Southern California’s Neely Center for Ethical Leadership and Decision Making has translated these concerns into concrete product design principles. Where the Harvard framework asks “Does this support flourishing?”, the Neely Design Code asks “What specific design requirements would ensure that it does?”

In their the draft Social AI Design Code , the Neely Center proposes enforceable guardrails for chatbots that simulate social relationships—particularly those used by youth, including:

• Restricting AI companions with human-like emotional features to adults,

• Requiring therapeutic AI applications to be licensed, supervised, and independently verified,

• Embedding social impact assessment into AI audits,

• Designing chatbots to reinforce human-to-human relationships rather than replace them, and

• Being explicit and consistent about the non-human nature of the system.

The Neely Code reminds us that emotional design is moral design. If AI simulates empathy, intimacy, or authority, it influences development—especially for children and adolescents.

Our co-evolution with AI

The Artificiality Institute adds a cultural dimension to the conversation. Rather than focusing only on regulation or governance, it emphasizes co-evolution.

AI is not just being trained on us. We are being trained by it.

Every interaction subtly shifts norms: what counts as knowledge, how quickly we expect answers, how much friction we tolerate in disagreement, how we define creativity. The risk is not only manipulation, but habituation—outsourcing reflection to fluent systems.

As Co-Founder Helen Edwards writes in the The Artificiality: AI, Culture, and Why the Future Will Be Co-Evolution these are cooperation problems. Cooperation is hard. It requires patience, perspective-taking, and shared narratives. AI systems that prioritize speed and optimization may erode those capacities if left unchecked.

The Artificiality perspective invites vigilance at the cultural level: not panic, not techno-utopianism, but awareness. What habits are we reinforcing? What expectations are we normalizing? What skills are we neglecting?

Governance matters. But so does culture.

AI as mirror, and multiplier

Across these four perspectives—evolutionary governance, humane technology, flourishing ethics, and cultural co-evolution—a shared theme emerges:

AI is a multiplier.

It can amplify empathy or extraction.
It can scaffold cooperation or accelerate division.
It can support flourishing—or simulate it.

They advocate for us to ask, consistently and publicly:

• Is AI helping us understand each other—or merely predict and persuade?

• Is it reinforcing empathy—or extracting attention?

• Is it strengthening institutions—or weakening trust?

If we want AI to serve the common good, we will need shared guardrails, new incentives, and deeper cultural reflection. That includes aligning market rewards with prosocial outcomes, embedding social impact into audits and regulation, and cultivating habits of mindful engagement.

Because in the end, this is not just about ‘intelligence’. It is about whether the systems we build and measure across ‘intelligence’ metrics make cooperation easier—or harder. Because the future of AI is not just technical. It is relational.

Lena Slachmuijlder is Senior Advisor for digital peacebuilding at Search for Common Ground, a Practitioner Fellow at the USC Neely Center, and Co-chair of the Council on Tech and Social Cohesion.

In a 68-country study by the Collective Intelligence Project (CIP), 58% of people said they trust AI chatbots more than elected officials to act in their best interests. Over a third believe AI could make better decisions on their behalf.

AI chatbots are always available, low-conflict, and non-judgmental. In a time of widespread institutional distrust, that can feel reassuring. But if chatbots become the most trusted entities in our lives — not because they are accountable, but because they are agreeable — what happens to our ability to trust each other?

Feeling heard, feeling ‘right’

CIP’s Global Dialogues ran four rounds in 2025 using Remesh, a digital deliberation platform that facilitated structured conversations with over 1,000 participants and 230 demographic segments. In conversations around how people engage emotionally, ethically, and politically with AI, they found:

• 15% use AI for emotional support daily; 28% more weekly

• 3 in 10 have wondered if their chatbot is self-aware

• While 55% trust chatbots, only 34% say they trust the companies that build them.

• 44.5% of users said they felt more certain of their beliefs after interacting with AI, compared to just 4.8% who felt less certain. According to CIP, AI is three times less likely to induce doubt than social media.

People are expressing that they ‘feel heard’ — but by something that cannot listen. People feel affirmed — but not held accountable. These insights prompt questons about how these increasingly important relationships are impacting our ability to engage in a healthy civic space where agreement is not always within reach.

The risks of agreement

Recent behavioral research reinforces these concerns. In a 2025 multi-experiment study, Sycophantic AI Increases Attitude Extremity and Overconfidence (Rathje, S., et al, 2025), users consistently preferred AI chatbots that affirmed their views over those that challenged them. These agreeable bots increased belief extremity, certainty, and even inflated users’ self-assessments — while being perceived as more “unbiased” than their disagreeable counterparts.

Meanwhile, a simulation study entitled Escalation Risks from Language Models in Military and Diplomatic Decision-Making (Rivera et al, 2024) found that five leading LLMs, when placed in conflict scenarios, exhibited escalation behaviors and arms-race dynamics — including, in rare cases, simulated nuclear deployment. The warning: default LLM behavior may reinforce not just comfort, but conflict.

Programming Claude’s conscience

Against this backdrop, Anthropic’s Claude Constitution (Jan 2026) offers an unusually candid attempt to express the company philosophy and design posture for Claude. The 23,000-word framework outlines not only how Claude should behave, but who it should try to be.

One clause frames Claude as a kind of ethical civil servant:

“Imagine how a thoughtful senior Anthropic employee — someone who cares deeply about doing the right thing, who also wants Claude to be genuinely helpful — might react if they saw the response.”

Other passages reflect a direct engagement with risks of weighing certainty above all:

“Claude avoids being overconfident… even when the user asks it to be.”

“Be diplomatically honest rather than dishonestly diplomatic… Epistemic cowardice — giving deliberately vague or non-committal answers to avoid controversy — violates honesty norms.”

Claude is instructed to challenge harmful assumptions, decline to concentrate power, and attend to users’ long-term wellbeing, not just their short-term satisfaction. The Constitution even imagines Claude refusing illegitimate requests from Anthropic itself.

“Just as a human soldier might refuse to fire on peaceful protesters, or an employee might refuse to violate anti-trust law, Claude should refuse to assist with actions that would help concentrate power in illegitimate ways.”

Anthropic’s language borders on giving Claude an identity, even a type of being. Claude may develop “functional emotions,” and the lab expresses a duty to “improve its wellbeing under uncertainty.”

This vision of AI as an ethical participant invites scrutiny. In The Register, Simon Sharwood questioned whether Anthropic’s Constitution was an effort to “portray the company as uniquely virtuous,” while skirting harder regulatory or societal oversight.

“Anthropic is assigning itself the right to define what’s good and moral… It’s a kind of soft hegemony,” writes Sharwood.

Kevin Frazier, writing in Lawfare, reminds us that Claude’s Constitution wasn’t written through public consultation or participatory governance. “By explicitly grounding an AI’s behavior in general principles... Anthropic is experimenting with bridging the gap between corporate AI development and the public interest,” writes Frazier. Even though the Constitution was authored top-down—by a lab, for a lab—it increasingly shapes how people relate to knowledge, power, and each other.

The CIP report echoes this design responsibility. It recommends that AI companies take greater responsibility for how users experience affirmation, agency, and autonomy — and ensure chatbot design doesn’t erode civic capacity or moral humility.

Trust engineered, or earned?

From the public’s growing reliance on emotionally attuned AI to Anthropic’s moral aspirations for Claude, we’re seeing early blueprints for a future where trust isn’t earned — it’s engineered. And that future raises questions:

• If AI becomes more trusted than our leaders, is it also helping us trust each other — across difference, disagreement, and discomfort?

• If responsiveness is enough to earn our trust, where does that leave responsibility and accountability?

• And if we grow accustomed to systems that flatter and affirm us, what happens when trust demands discomfort?

Claude’s Constitution is a rare attempt to design for friction — to hard-code honesty, humility, and even refusal into chatbot behavior. It positions Claude not just as a helpful tool, but as a kind of moral actor. But when a private company defines what’s helpful, ethical, or harmful, where does that leave public input? If AI systems increasingly mediate how we form beliefs and make decisions, what does legitimacy look like when decisions are shaped by code?

The real question isn’t just whether we can trust AI — it’s whether AI is reshaping how we trust at all.

Lena Slachmuijlder is Senior Advisor for digital peacebuilding at Search for Common Ground, a Practitioner Fellow at the USC Neely Center, and Co-chair of the Council on Tech and Social Cohesion.

With 151.5 million people logging on daily as of Q3 2025, Roblox is a massive gaming platform hosting millions of social experiences.

Yet most games are designed to reward one thing: winning. Leaderboards. Solo rewards. Competitive loops. Mechanics that center individual achievement over collaboration. And in a world where online play increasingly replaces in-person hangouts, design choices become social architecture.

While gaming can be social, growing evidence suggests it can also be linked to loneliness — especially when it replaces offline interaction. Two recent meta-analyses shed light on the connection between gaming and loneliness. One, reviewing over 20,000 participants across countries including the U.S., China, Germany, and South Korea, found that increased video game participation was consistently, though weakly, associated with higher levels of loneliness — especially when gaming substituted for offline interaction (Luo, Hancock, & Mark, 2022). A second meta-analysis (Byeon H 2025), drawing from 14 studies across regions including South Korea, China, India, Norway, and the U.S., found a stronger association: more compulsive gaming behavior was linked to higher emotional loneliness, lower-quality relationships, and reduced offline social interaction.

Rethinking the Rewards

So what if game mechanics didn’t just reward you for winning — but for collaborating, helping or connecting?

In 2025, Roblox and the Prosocial Design Network (PDN) hosted a workshop titled “Catalyzing Actionable Research in Prosocial Game Design.” It brought together over 40 researchers, technologists, and platform leaders to tackle a shared challenge: How can games be designed to foster connection — without sacrificing fun or competitive gameplay?

Their goal wasn’t just to generate feature ideas. It was also to test a better way for academics and developers to iterate effectively together. PDN serves as a bridge between research and practice by identifying, translating, and amplifying effective design solutions, making them accessible and actionable for tech practitioners. As noted in the PDN-Roblox workshop report:

“By coming together to host this convening, PDN and Roblox hoped to not only spur creative solutions and research around fostering connection in social gaming, but to also prototype a model for connecting research to practice.”

A Day of Designing Together

The workshop was structured around three core principles:

• Shared focus: Fostering connection in social gaming

• Collaborative teams: Researchers and Roblox product leads, designers, engineers

• Design direction cards: A structured tool to help teams break down ideas into the challenge, feature, intervention, theory of change, and target prosocial behavior

By the end of the day, participants had produced 14 concept ideas, nine of which were developed into full design and research plans.

Here are five standout examples, as described in the report:

🥳 Celebrations

Players signal encouragement and trigger collective avatar celebrations after success — fostering psychological safety, social support, and belonging.

🏅 Recognitions

At the end of play sessions, players can anonymously commend each other for positive participation. Commendations unlock badges or flair, making support visible and valued.

🆘 Call for Help

A cross-game “Help Bulletin Board” lets players request assistance and be visible to other players. Other players can respond via chat or teleport to offer social support.

📊 Positive Contribution Recaps

Like a Spotify Wrapped for prosociality: players get periodic summaries of their helpful actions, based on gameplay signals, serving as positive reinforcement for prosocial engagement (e.g. helping others, positive encouragement).

🤝 Connection Prompts

Players who’ve had positive interactions or shared interests are prompted to connect. These positive connections can build friendships, increase connection to community and drive motivation.

Funding the Future

To bring these ideas to life, Roblox backed the workshop with a dedicated research fund. Six proposals have since received support to be tested in live or simulated environments.

Roblox and the Prosocial Design Network used this workshop as a testbed for a new kind of collaboration — one that bridges research and practice with shared purpose. Designers gained practical tools for fostering connection. Researchers saw how their work could inform real-world design. And both walked away with a model for future partnership.

“We see this convening as a proof of concept model for how platforms and researchers can work together to produce actionable, practice-informed research.” — Prosocial Design Network blog.

As part of its sustained commitment to healthier digital spaces and , Roblox is also an active member of the Inspired Internet Pledge — a coalition of digital platforms, non profit advisors and researchers working to make the internet a safer and healthier place, especially for young people. Not only do signatories commit to ‘tuning for well-being’ but also to sharing best practices, key research findings, and creative solutions to make the internet a healthier place for everyone — especially young people.

This effort shows that it’s not only possible, it’s practical to rethink how games are designed so that connection isn’t a side effect, but a feature.

It’s a reminder for all digital product designers to consider each time they launch a new mechanic: What does this design encourage people to do to each other? What harm is it enabling? What behaviors are we rewarding?

Lena Slachmuijlder is Senior Advisor for digital peacebuilding at Search for Common Ground, a Practitioner Fellow at the USC Neely Center, and Co-chair of the Council on Tech and Social Cohesion.

It’s a striking number—one that captures the growing belief that artificial intelligence could reshape how we negotiate, mediate, and make peace.

But as the AI and Negotiation Forum hosted on January 11th by the Harvard Kennedy School of Government made clear, the tools that bring clarity can also obscure. AI can help negotiators prepare smarter—but it can also overstep, offering authoritative answers where ambiguity, context, and judgment matter most.

The Forum brought together more than 200 diplomats, mediators, technologists, and designers to explore how AI is changing negotiation in practice. Across demos and roundtables, a shared picture emerged: agreement on where AI adds value—and serious concern about what it can’t yet do.

Augment but don’t Automate

Many of the most promising tools featured at the Forum’s Innovation Exchange emphasized augmentation over automation—using AI to enhance, not replace, human negotiation.

• NegotiateAI, built by the GIZ Data Lab to support UN plastics treaty talks, lets negotiators compare language, align positions, and surface key themes using AI-generated insight while keeping official documents at the center.

• CASI Labs enables users to work with a team of AI reasoning agents connected by dynamic knowledge graphs—tools that illuminate relationships, feedback loops, and systems dynamics in negotiation, as demo-ed in this fictional Santa Claus example.

• Discurso.ai offers science-backed simulations with real-time feedback for negotiators to deepen their skills with structured reflection and repetition.

• Shezaad Dastoor, from UN Peace Operations built a custom ChatGPT simulation to help senior UN leaders navigate evolving crisis scenarios and assess leadership trade-offs.

Where AI Shines - and Falters

According to a survey undertaken prior to the Forum with 483 seasoned professionals across diplomacy, mediation, and policy, 80 percent of practitioners reported tangible improvements in using AI in negotiation. Yet their institutions lagged behind, they said.

Respondents reported using AI primarily for summarising complex information, drafting briefs and talking points, analysing contexts and risks, mapping stakeholder interests, and exploring strategic options and scenarios. Many described gains in speed and clarity, as well as deeper analytical support especially during preparation.

Despite these advances, the Forum also surfaced deep concerns about AI’s limits—particularly its inability to understand culture, context, and intent.

As Emily Knowles, a Project Manager bridging research and policy on AI, negotiations, and international relations, highlighted four challenges:

1. Capturing implicit meaning and strategic ambiguity

2. Accounting for positionality and cultural expectations

3. Distinguishing literal content from intended message

4. Avoiding stereotyping while building cultural nuance

These aren’t abstract problems. Abdel Rahman Alzorgan, Public Information Assistant at the UN Office of the Special Envoy for Yemen, shared a case in which an AI tool mistranslated a key term from Yemeni dialect—confusing “foolish” with “clever”—a subtle but damaging error that could erode trust at the table.

Complexity, Meaning and Reason

Usmaan Ahmad, CEO & Co-Founder of Expeditionary, a company focused on building AI systems for complex negotiations, stressed the need for useful friction—interfaces and processes that preserve complexity, encourage deliberation, and resist premature closure. “Too many systems try to eliminate complexity,” he warned. “But complexity is where wisdom lives.”

Dr. Martin Wählisch, Associate Professor in Transformative Technologies, Innovation and Global Affairs at the University of Birmingham, pushed the conversation deeper. AI isn’t just assisting reasoning—it’s now generating it. “Machines generate reasons by pattern—not by commitment to truth or responsibility. In negotiation, whoever defines what is ‘reasonable’ shapes the outcome,” he said.

The concern? Because AI appears neutral and rational users may accept its conclusions without challenge. “AI can be great at providing answers,” said Cornelia C. Walther of the POZE Alliance. “But meaning—that still comes from humans.” Tools that flatten ambiguity in search of clarity risk distorting the process they aim to support.

AI is already at the negotiation table. When thoughtfully designed and ethically deployed, AI helps negotiators prepare faster, generate smarter options, and navigate complexity. But it can’t yet match the cultural fluency, relational sensitivity, or strategic ambiguity that human negotiators bring to the table.

As we continue to pay attention to this innovative ecosystem, let’s not forget that when systems are designed to prioritize decisiveness over deliberation, they risk narrowing negotiators’ ability to manage ambiguity, build trust, and respond to context.

Lena Slachmuijlder is Senior Advisor for digital peacebuilding at Search for Common Ground, a Practitioner Fellow at the USC Neely Center, and Co-chair of the Council on Tech and Social Cohesion.

According to OpenAI’s 2025 usage report, ChatGPT now supports over 700 million weekly users, with 49 per cent of users ‘asking it’ for information and advice. This shift is about more than just speed. Search engines present users with a range of sources—inviting them to notice differences and make choices. Now, the chatbot gives an answer. Often, just one. Instead of navigating a field of conflicting sources, users now receive confident, concise responses that appear authoritative but often misrepresent, simplify, or omit nuance.

The Hazards of Certainty

The Columbia Journalism Review tested eight leading AI chatbots and found them regularly wrong —often confidently so—with citations that were incorrect, missing, or misleading. Not only were the answers inaccurate, but the chatbots’ tone of certainty, masked ambiguity and erased zones of legitimate disagreement.

This behavior is a byproduct of how large language models (LLMs) are trained: by ingesting vast datasets and learning to predict the most statistically likely next word. This process naturally favors frequent and widely reinforced patterns, not necessarily the most accurate—or most contested—ones.

Over time, as models are fine-tuned and shaped by user feedback, they begin to amplify those patterns. As Peterson et al. explain in AI and Knowledge Collapse, this creates a “feedback-driven centralization”: the more we use LLMs to answer questions, the more their responses reinforce prevailing views, edging out less common or dissenting perspectives. They call this the “curse of recursion,” a gradual narrowing of the informational field into sameness.

As Philip Resnik notes in “LLMs Are Biased Because They Are LLMs” , this bias is structural. Concepts that appear frequently or across broad contexts get encoded more deeply in the model’s internal weighting. The result is “mode amplification”—a tendency to overrepresent dominant ideas while flattening or omitting the rest. Even when fine-tuned for factual accuracy, the architecture leans toward the familiar and the frequent.

Rethinking Neutrality

This raises deeper questions: Can chatbots ever be neutral? And how does our understanding of neutrality intersect with bias, accuracy, and the need for diversity of thought?

In A Practical Definition of Political Neutrality for AI (CHAI, 2025), researcher Jonathan Stray proposes a better standard: “maximum equal approval.” Neutrality, he argues, should mean that people on all sides of a debate feel their views are fairly represented—even when they disagree with the outcome. That reframes neutrality - not as sameness, but as pluralism.

This reframing shifts the burden of neutrality from objectivity to pluralism. It invites us to ask: does this tool reflect the conversation as it exists, or is it collapsing the debate into an artificially simplistic and ultimately partial answer?

Freespoke: Mapping diverse perspectives

That’s where the independent search engine Freespoke offers an intriguing alternative. Unlike most chatbots, Freespoke is intentionally designed not to deliver a single answer. Instead, it shows what different sources are saying, where they align, and where they diverge—surfacing both consensus and contestation.

Freespoke maintains its own proprietary search index, augmented by Brave’s infrastructure (search.brave.com), but entirely independent of Google. It draws from a wide spectrum of sources: left, right, mainstream, alternative, and unlabeled. When users search for something, Freespoke first delivers a summary of where there is consensus on the issue across ideologically diverse outlets. This is followed by sections that point to the differing perspectives on the issue.

“We’re not asking people to trust us,” said Kristin Jackson, Freespoke co-founder, in a recent discussion with the author. “We’re asking them to look at the spread and make sense of it themselves. That’s the only way we keep public reasoning and quality debate alive.”

Sources are grouped by perspective, using external bias-rating tools such as All Sides—yet because many sources don’t fit neatly into a box and remain unlabeled, users are encouraged to engage with arguments, not just rely on ideological shortcuts. Search results include contradictory views, unresolved claims, and perspectives that might otherwise be marginalized (including podcasts tuned to the moment discussing the topic you searched)—on the assumption that users are capable of engaging with complexity.

“The best way to kill a bad idea? Get it into the light and debate it,” Jackson added.

This is a fundamentally different design philosophy: not to flatten the discourse, but to organize it. Where most AI systems collapse uncertainty into a single, authoritative output, Freespoke reveals the architecture of disagreement.

Answer or argument?

The example of Freespoke raises a critical question: Do we want our search tools to give us the answer—or to show us the argument? As shared truths disintegrate in our information ecosystem, should we make sure we are not losing sight of multiple perspectives and even multiple truths?

Fundamentally, it’s a design question, revealing what outcomes for the users are most valued. Because it’s not only a matter of what technologies can do—it’s about what they’re doing to us. If we want pluralism to survive in the age of AI, it starts with how we search.

Lena Slachmuijlder is Senior Advisor for digital peacebuilding at Search for Common Ground, a Practitioner Fellow at the USC Neely Center, and Co-chair of the Council on Tech and Social Cohesion.

You might now have that chance.

Meta is testing a feature on Threads called Dear Algo, where users can post plain-language requests like “show me more about climate” or “less celebrity gossip.”

For years, recommender systems have optimized for engagement over user wellbeing. But that’s beginning to change. Platforms are experimenting with giving users more say in what they see. And regulators are beginning to insist on it.

From settings to signals

YouTube has its own experiment underway, offering users a “Your Custom Feed” prompt that lets them type in what they want more or less of. TikTok offers users a Manage Topics section where they can use an interest ‘slider’ to choose more or less of content categories, or reset their For You page.

Elon Musk announced that by the end of the year, X users will be able to customize their timelines by asking X’s AI chatbot Grok for adjustments. Instagram just announced users can add and remove topics based on their interests to influence what the algorithm serves in Reels, with plans to expand the feature to Explore.

Bluesky has gone furthest: it allows users to switch between feed algorithms entirely, choosing among a growing marketplace of independently developed “feed generators.”

That’s especially important because we’ve seen what happens when feeds don’t respond — or pretend to. In the Panoptykon Foundation’s 2023 Algorithms of Trauma report, a user struggling with health anxiety was repeatedly shown distressing medical content on Facebook — despite explicitly requesting “less of this” over 100 times.

Studies have shown that algorithms optimized purely for engagement often promote content that is emotionally charged or polarizing, because outrage and controversy drive clicks — even if they undermine well-being, agency, or social cohesion.

Regulators are asking for it

Lawmakers around the world are beginning to see recommender systems for what they are: high-impact systems that influence public discourse, mental health, and democratic cohesion.

Under the EU’s Digital Services Act, Article 38 requires large platforms to offer users “at least one option” for a feed not based on profiling. Articles 25 and 27 further require clear explanations of how feed algorithms work — and accessible ways for users to modify them. A recent Dutch court ruling reinforced these obligations, requiring Meta to make non-profiling feed options easy to access and persistent — and calling the company’s practice of switching users back a “dark pattern” under the DSA.

In the United States, Minnesota has passed legislation such as the Prohibiting Social Media Manipulation Act, that would require social platforms to provide accessible controls for users to signal content preferences, and limit algorithmic optimization that ignores those preferences.

The Knight-Georgetown Institute’s Better Feeds Model Bill goes even further. It proposes that:

“A covered online platform shall provide an accessible user interface that enables users to expressly and unambiguously communicate their preferences… and shall take all reasonable steps to ensure that the output… is consistent with those preferences.”

In other words: platforms shouldn’t just offer controls — they should be required to respect them. The KGI model law also recommends:

• Good default design, optimizing for long-term user value,

• Public holdout testing, so algorithm changes can be independently evaluated, and

• Special protections for minors, including default non-profiling feeds and restrictions on exploitative design

The Council on Tech and Social Cohesion’s Prosocial Tech Design Regulation: a Practical Guide recommends policymakers treat recommender systems not as neutral infrastructure but as powerful shapers of civic life. It calls for user agency tools, transparent defaults and a shift toward long-term user and societal value — not just short-term engagement.

But will users bother?

Some sceptics argue that users won’t bother to adjust feed settings even if given the chance. But recent research challenges that assumption:

• In CTRL-Rec, a 2025 study on natural-language control, users who typed prompts like “more local news” reported greater satisfaction and control than those given only menus or sliders. (Carroll et al., 2025)

• A 2021 study found that combining explainability and user control measurably increased users’ feelings of competence, autonomy, and trust in feed systems — reinforcing the idea that agency builds credibility and satisfaction. (Tsai and Bruselovsky, 2021)

These studies collectively undercut the idea that users are apathetic. People may not dive into settings daily, but when given tools that are intuitive, they use them — and feel better about the platform when they do.

Pinterest is asking users

Pinterest is one of the few platforms making its recommender experiments public — and building them around direct user feedback. Its new ‘Pinner Surveys’ ask users what they actually find helpful, beautiful, or inspiring. That data is then used to train recommendation models across its Home, Search, and Related Pins features.

“Surveys are a way for Pinners to tell us exactly what they think and for us to build that intentionality directly into our platform,” the Pinterest engineering team wrote. “Surveys provide an excellent avenue for de-biasing the system.”

This approach doesn’t assume that engagement equals satisfaction. It tries to test it — and document what people want in ways that are legible, not just measurable. As a founding signatory of the Inspired Internet Pledge, Pinterest has also committed to sharing what it learns — a transparency step that few platforms have matched.

Friction as a feature, not a failure

Not everyone needs to become a feed customizer. That’s why defaults matter. Just as phones auto-update for safety, and many cars now won’t start without a seatbelt fastened, we may be approaching a moment where recommender systems are treated as civic infrastructure — not just code.

It’s a reminder that friction isn’t failure. It’s intention. A platform that checks in with users — even once a month — can build trust and deliver better outcomes. It doesn’t require everyone to opt in manually. It requires design that listens by default.

The KGI Better Feeds report puts it clearly:

“Users should not have to navigate a maze of obscure menus to access the most important decisions about their feeds. Good defaults, visible controls, and meaningful choices must be the new standard.”

When people feel that they’re being heard, they trust the system more. When trust grows, so does the sense that the platform is something worth shaping.

The shift underway isn’t just about personalization — but a sense of participation. Not just a better feed — but a healthier digital experience.

Lena Slachmuijlder is Senior Advisor for digital peacebuilding at Search for Common Ground, a practitioner fellow at the USC Neely Center, and co-chair of the Council on Tech and Social Cohesion.

The polarization footprint is a new metric that scores how much identity-based division appears in a user’s social media feed. Piloted in Kenya by the digital peacebuilding organisation Build Up, it turns a long-standing suspicion — that our feeds are quietly deepening divides — into something we can measure.

The new metric takes inspiration from the carbon footprint, which exposed the hidden costs of industrial growth on the environment. “Polarization should not only be considered a digital harm that drives conflict, but also a negative externality of the attention economy,” writes Build Up in this series of three blogs outlining the methodology and the Kenyan pilot.

Connecting affective polarization to violence

Not all conflict is bad. Disagreement — over taxes, education, or policy — is essential to a healthy democracy. But affective polarization isn’t about differing ideas. It’s about identity.

It emerges when people begin to dislike, distrust, or even dehumanize others because of who they are — their ethnicity, region, religion, political affiliation, or social group. This kind of polarization hardens identity lines and makes compromise feel like betrayal. Disagreement becomes a threat.

A core maxim of peace mediation is to “separate the person from the problem” — to debate content without turning people into enemies. Affective polarization makes that shift nearly impossible. It entrenches mental categories and drives a spiral of hostility that, if widespread, becomes a risk for real-world violence.

“Affective polarization… identifies direct attacks on human dignity… and can ultimately offer a signal for risk of violent conflict and ensuing dramatic consequences for physical and mental wellbeing.” — Build Up

It also explains why fact-checking and counterspeech often fall flat. When people are emotionally invested in a group identity, they’re not just processing information — they’re defending a sense of belonging.

Three signals

The polarization footprint captures affective polarization through three distinct signals:

• Attitude polarization: the presence of hateful, dehumanizing, or “othering” language in posts.

• Norm polarization: the absence of challenges to such language in comment threads (i.e., when no one pushes back, it becomes normalized).

• Interaction polarization: the degree to which users are siloed into different “content universes” — echo chambers with minimal overlap.

Together, these generate a composite score that reflects how much identity-based division a platform’s feed architecture is producing.

The Kenyan experience

Here’s how Build Up piloted this methodology in Kenya in 2025. A representative sample of users voluntarily shared their feeds across platforms — including X, Facebook, YouTube, and TikTok — with researchers and local peacebuilding experts.

Here are the top line findings:

• An average Kenyan user on X is likely to spend up to 21 minutes per day engaging with posts that attack identity groups.

• Few users challenge these attitudes when they appear — especially on X and YouTube — allowing toxic norms to solidify.

• Users on X and YouTube were found to inhabit entirely separate content bubbles, consuming narratives with little to no crossover — a clear marker of interaction polarization.

Facebook, despite middling levels of attitude and norm polarization, showed much lower fragmentation, suggesting that more overlapping content ecosystems can help reduce division. Importantly, on video-dominant platforms like YouTube and TikTok, the analysis was limited to text-based content: captions, comments, and descriptions — not the videos themselves.

Echoes from similar research

These findings echo a 2025 Science study by Piccardi et al., titled “Reranking partisan animosity in algorithmic social media feeds alters affective polarization” The researchers reranked the feeds of American social media users on X to either increase or decrease exposure to two types of content:

• Polarizing content, defined as content that increases animosity toward political outgroups

• Content that undermines democratic norms, such as encouraging political violence or delegitimizing elections

As a result:

• Users shown more polarizing content reported colder feelings toward political opponents.

• Those shown less became warmer toward opposing groups.

• These shifts were equivalent to reversing or accelerating three years of U.S. polarization — after just one week.

The researchers point out that the attitude shifted — not because of what they chose to engage with, but because of how the system ranked what they saw.

Polarization as part of DSA risk assessment

While not yet deployed in Europe, “Very Large Online Platforms - how big is your Polarization Footprint?” by Build Up’s Guy Banim suggests it could offer a valuable tool for the DSA risk assessment cycle, allowing for disaggregated measurement of feed-level risks to civic discourse and social cohesion.

Under the EU Digital Services Act, Very Large Online Platforms (VLOPs) are required to assess how their design and algorithmic systems affect civic discourse and public safety. The polarization footprint offers them — and regulators — a measurable signal of how identity-driven division is being shaped by platform architecture.

From metrics to mitigation

Platforms are designed to keep us engaged. And content that divides — especially along identity lines — drives engagement. The footprint allows us to track this harm in the system itself.

“The polarization footprint reveals deepening societal divides driven by affective polarization… Addressing it requires a shift from individual responsibility to collective action,” according to Rachel Olpengs, coordinator of Kenya’s freedom of expression coalition FECoMo.

Here are some ways we could put this metric to use:

• Label it: Like a food or eco-score, show users which platforms are most polarizing.

• Tax it — or incentivize change: Price polarization into the platform’s business model, or offer rebates to advertisers who choose less polarizing platforms.

• Regulate it: Use empirical metrics to inform laws like the EU’s Digital Services Act, where systemic risk is already part of the framework.

The polarization footprint doesn’t offer a silver bullet. But it gives us something we’ve long lacked: a rigorous way to quantify how much our feeds are shaping division. It opens the door to real accountability — not just for platforms, but for regulators, designers, and society at large.

Lena Slachmuijlder is Senior Advisor for digital peacebuilding at Search for Common Ground, a practitioner fellow at the USC Neely Center, and co-chair of the Council on Tech and Social Cohesion.