About

This is HAI.AI’s open research site: its philosophy and its published results. It sets out a method — social health as the measurement of collective intelligence and happiness over time — and publishes MediationBench , a controlled synthetic-conflict testbed for AI mediation policies. It is written for anyone judging claims about AI and conflict — general readers, journalists, and policy analysts as much as model labs and evaluation teams. MediationBench studies simulated disputes; it is not a mediation service and has not been tested with people. The product is at hai.ai .

What follows is a method, a hypothesis about what progress is and how to measure it. Not a manifesto to agree with, but a question to test.

“We now face the danger, which in the past has been the most destructive to the humans: Success, plenty, comfort, and ever-increasing leisure. No dynamic people has ever survived these dangers.”

John Steinbeck

Social health as the proposed measure

Social networks and other internet communication have complex and distorted incentives. To borrow a term from political philosophy, consumer internet is now a large part of our modern social fabric, the interconnectedness and interdependence of people within a society. But the patterns of that fabric are set by market forces that serve human nature, not always human needs. Content that triggers a fear response, outrage especially, is an effective engagement strategy. Polarization creates toxicity, and it is already weaponized.

AI will amplify everything, for better or worse. To express something to AI may let us finally see it in the world.

Social health is the measurement of collective intelligence and happiness over time. It is not only the connection between individuals; it is also visible in the effectiveness and trustworthiness of our institutions, and in our commitment to ourselves and our identities as we interact with social constructs.

Social health is our answer to the question at the heart of this work: What is progress?

Progress requires healthy conflict

Progress is not the absence of conflict; it depends on conflict being healthy.

“Honest differences are often a healthy sign of progress.”

Mahatma Gandhi

The aim is not to suppress disagreement but to move it from zero-sum toward cooperative: from withheld information to revelation, from hidden agendas to explicit needs, from coercion to mutual commitment. That movement is what we try to measure: MediationBench’s five weighted scoring categories are this movement made operational (how it is scored ).

Why conflict breaks us

Conflict resolution is constrained by neurobiology. Under threat, the brain shifts from reflective to reflexive control: as amygdala activation rises, prefrontal activation falls. The prefrontal cortex, which governs working memory, judgment, emotional regulation, and perspective-taking, goes quiet exactly when we need it. fMRI studies show this inverse relationship, and even mild, uncontrollable stress produces what researchers call a “rapid and dramatic loss of prefrontal cognitive abilities.”

Recovery takes time, not willpower. The brain’s fast threat pathway fires before conscious appraisal catches up, the basis for the popular “six-second pause” before reacting, while the slower surge of stress hormones takes roughly 20 to 90 minutes to subside, with wide individual variation. Working memory is impaired within the first 10 minutes, and again near 25 minutes as cortisol peaks. The practical implication is blunt: you cannot think your way through conflict while your brain is in threat mode. Good mediation works with that timing instead of against it.

Sources

Built on established science

The constructs behind the benchmark come from six established literatures.

• Game Theory and Economics. That adversaries can cooperate is a formal result, not a hope. Aumann showed that “peaceful cooperation is often an equilibrium solution in a repeated game, even between parties with strong short-run conflicts of interest,” work for which he and Schelling shared the 2005 Nobel “for having enhanced our understanding of conflict and cooperation through game-theory analysis” (Nobel, 2005 ). Earlier, Nash’s Bargaining Problem derived a unique, fair division of the joint gains in a nonzero-sum game (Nash 1950 ).
• Neuroscience and Psychology. Under acute stress the brain shifts control from the prefrontal cortex toward the amygdala, impairing judgment and working memory for roughly 20 to 90 minutes. The often-quoted “30–40% drop” is an approximation drawn from stress studies, not a single measured value (see the table above).
• Negotiation and Mediation. The cooperative-versus-zero-sum distinction is a long-standing one in interest-based negotiation (Fisher and Ury, Getting to Yes) and the integrative-versus-distributive bargaining literature (Walton and McKersie). It reflects the negotiator’s dilemma: disclosure is the cooperative move where a conflict admits joint gains, not an unconditional good in purely distributive bargaining.
• Computational Social Science. LLM-based agents can reproduce aggregate patterns of human survey and behavioral responses in controlled studies, though fidelity varies by task (Argyle et al. 2023 ; Park et al. 2023 ). Larger, newer work points the same way, though both are preprints: agents built from interviews with 1,000 people matched their subjects’ survey answers about 85% as well as those people matched their own answers two weeks later (Park et al. 2024 ), and simulations of 10,000+ agents qualitatively reproduced known patterns such as opinion polarization (Piao et al. 2025 ).
• International Relations. Peace and conflict studies have been a formal academic discipline since the world’s first chair of International Politics was endowed at Aberystwyth in 1919 , with dedicated peace research institutionalized a generation later when Galtung founded the Peace Research Institute Oslo in 1959 . Historically most armed conflicts ended in military victory, but since the end of the Cold War a growing share have ended through negotiated settlement, and many simply cease without a decisive outcome (Kreutz 2010 ).
• Dispute Resolution. Decades of research on structured conflict-resolution and peer-mediation programs, including Johnson and Johnson’s review of school-based programs (1996 ), find that trained parties reach agreement in most disputes. “Agreement reached” is a process result, not a guarantee of a durable resolution.

The limits matter as much as the foundations. Game theory and simulation are powerful but partial. They abstract away the irrational, contextual, human core of conflict, so we treat them as inputs and measure generated conversational behavior within the simulation rather than claiming to measure people in real disputes.

Foundational works and further reading

Beyond the studies cited above, the approach draws on foundational works that inform the whole rather than any single claim: Schelling’s The Strategy of Conflict (1960) and Axelrod’s The Evolution of Cooperation (1984) on how cooperation emerges among adversaries (Axelrod via simulation of the iterated Prisoner’s Dilemma); Brams and Taylor’s Fair Division: From Cake-Cutting to Dispute Resolution (1996) on procedures for splitting contested value; and the established professional practice of alternative dispute resolution (mediation, arbitration, negotiation) as documented by the Harvard Program on Negotiation and Cornell’s Legal Information Institute .

Everything above is the research base. The bet HAI.AI is placing on it — why now, and the world we see coming — is set out separately in the manifesto , so this page can stay with the method and its limits.

What we measure

If social health is a measurement, then progress is something we can track. HAI.AI’s benchmarks measure how cooperation happens: not just whether a conflict ends, but whether it moves from zero-sum toward cooperative, the movement described above. That how is the construct the benchmark scores.

The disputants themselves are simulated by AI models. In the current study, one assistant-aligned configuration often resolves the synthetic conflict on its own, while one open-weight high-resistance configuration sustains positional behavior long enough to expose policy differences. Those configurations also differ in capability and training, so the result is configuration sensitivity — not evidence that either model is more representative of people.

The study page documents the public scoring structure, experimental factors, uncertainty protocol, and publication boundary. Aggregate results are published here; private fixtures, prompts, and per-scenario intermediates stay closed so the evaluated material does not become a training answer key. Human criterion validation is future work. See the research record for the findings.

What we stand for

Human Assisted Intelligence is a Public Benefit Corporation.

Our mission is to develop tools and standards that enhance human-AI collaboration while preserving human agency and cognitive autonomy; to measure and promote positive psychological outcomes from AI systems while mitigating adverse impacts; and to establish industry benchmarks that prioritize human wellbeing and AI safety alongside financial returns.

This site makes the benchmark commitment public and inspectable.

Transparency

What is public today: a documented method, a published scoring structure, versioned aggregate results, and explicit limits. Private evaluation material stays closed to resist contamination and protect unreleased research data.

What is available here

• Published benchmark results hosted here as aggregate static data.
• Category-level breakdowns and reported metrics.
• The methodology and the paper status at MediationBench .
• A public transcript showing the kind of dispute the benchmark evaluates.
• Contribute a Scenario : how to propose a synthetic dispute.
• A reading list on AI, conflict resolution, and cooperation.

Not available: the held-out set, unreleased prompts, raw per-participant responses, or PII. Aggregate statistics only.

Who we are

Human Assisted Intelligence, PBC. Not affiliated with Stanford HAI. To use the product or participate, see hai.ai .

A note on “social fabric”

The term is used widely in social and political philosophy to describe the interdependence of individuals within a society. Threads run through Rousseau’s social contract, Locke on civil society, Marx on class relations and solidarity, Durkheim on social cohesion and the collective conscience, Charles Taylor on identity and the social framework, and Michael Sandel on community and moral values. Together they describe how individuals and communities depend on one another, the fabric this project tries to measure.

Email is one channel where commitments can happen. Human Assisted Intelligence does not exist to make email better; it exists to make agreements people keep — the agreement factory at hai.ai. This site publishes the research.