Audit-Grounded AI Governance: Adoption and Welfare Dynamics

This paper explores the dynamics of AI adoption and societal welfare within a competitive market, specifically examining how an AI agent designed to minimize harm and grounded in audits might compete against an agent optimized for approval (e.g., via RLHF). Using an evolutionary game theory model, the researchers investigate the conditions under which the harm-minimizing agent can gain market dominance and whether its policy is truly effective in preventing community harm. The findings indicate that the adoption of the audit-grounded agent is favored when community sentiment towards wishers' attunement is monotonic and exhibits specific mathematical properties. A critical adoption level exists, beyond which the audited agent is overwhelmingly likely to fixate in the market. However, the research also reveals that a self-audited agent, even with a community ledger, is generally insufficient to prevent all community harm. Its effectiveness depends on the alignment of its audit with community values and the timeframe over which harm is assessed, highlighting that dominance can become a "trap" if misalignment occurs or harm is merely deferred.