Beyond mediation: an evolutionary benchmark for emotionally and normatively competent AI

Area of Science:

• Artificial Intelligence
• Computational Social Science
• Game Theory

Background:

• Current AI evaluation benchmarks often focus on logical inference or computational efficiency.
• Effective AI mediation in social and legal contexts requires assessing adaptive and emotional capabilities.
• Achieving fairness and stability in dispute resolution is complex and may not align with simply being 'right'.

Purpose of the Study:

• To present an agent-based simulation framework for evaluating AI mediators.
• To measure key dispute resolution traits: emotional regulation, compromise, and trust repair.
• To compare AI mediator policies across diverse negotiation tasks and define explicit trade-offs.

Main Methods:

• Developed five canonical negotiation tasks (S1, H1, H2, H3, C1) with varying psychological reactivity and asymmetry.
• Employed Monte Carlo simulations to compare five distinct AI mediator policies.
• Evaluated policies based on agreement probability, time to agreement, outcome inequality, and task robustness.

Main Results:

• Stable trade-offs exist: strong AI control reduces negotiation time but increases inequality.
• Transparent, weak AI interventions better balance efficiency and fairness across different scenarios.
• Policies favoring cross-task robustness were favored by evolutionary selection over single-metric optimization.

Conclusions:

• AI mediation evaluation should prioritize process control, incorporating emotion and norms.
• The developed benchmark redefines AI assessment for mediation, moving beyond pure efficiency.
• Future work (v2) will incorporate more complex emotional dynamics and event-driven shocks for enhanced realism.