Open benchmark results from HAI.AI
One answer is social: getting better at resolving conflict. We test whether AI mediators can move a simulated dispute — both sides played by AI models, not real people — from zero-sum standoff toward workable agreement, and publish the results here.
What we measure
HAI.AI benchmarks ask whether a model or mediator changes the process of a simulated conflict: more disclosure, clearer needs, reciprocal commitments, and less zero-sum behavior.
Model-played participants begin defensive and emotionally activated, then continue the seeded conversation. Open-source, minimally moderated participant models keep the conflict realistic — assistant-tuned models tend to de-escalate on their own.
Runs are labeled by suite version, participant model, judge model, mediator type, and mediator model.
Scores are aggregate, model-judged results under defined conditions, not guarantees of safety.
For foundation model teams
The benchmark is meant to be useful to labs that want to understand how their models handle adversarial, emotionally loaded conversations. We are especially interested in feedback on scenario realism, variance, human review, and the metadata needed for reproducible comparisons.
A concrete example
The sample shows the non-held-out SDK/free benchmark scenario: two AI-played business partners start from accusation, fear, and positional bargaining. It is a safe excerpt, not the hidden evaluation set.
Read the excerpt"I am not pretending this is an equal split after I kept this place alive while Raj checked out." Maya, public sample scenario
Human Assisted Intelligence is a Public Benefit Corporation
We want AI that makes us better at being human, and the benchmarks to prove it under careful, limited claims.
The product lives at hai.ai; this site publishes the philosophy and the numbers.