Behavior
Track what agents do under resource pressure, partial observability, and changing social value orientation.
Govern AI systems by mapping behavior, harm, evidence, and control points.
EthicaAI turns multi-agent reinforcement learning experiments into an audit-friendly vocabulary for AI ethics risk governance. The focus is not a generic checklist. It is a structured way to ask which incentives produce cooperation, failure, concentration of welfare, or policy drift.
Harm
Connect model behavior to stakeholder impact: welfare loss, unfair allocation, manipulation, and brittle coordination.
Controls
Translate findings into review gates, monitoring criteria, red-team scenarios, and escalation thresholds.
Governance questions this page helps answer
- Which incentives make cooperation stable instead of temporary?
- Where does an AI policy improve average welfare while harming a minority case?
- What evidence is strong enough for an internal model risk review?
- Which conditions should trigger a human review before deployment?
Evidence path
Start with the 70-figure research gallery, then use the compliance evaluation guide to convert research evidence into audit criteria. For external citation, use the Zenodo DOI and repository linked from the EthicaAI home page.