Evolution of fairness in hybrid populations with specialised AI agents

A new research paper from Zhao Song, Theodor Cimpeanu, and Chen Shen, published on arXiv (ID: 2602.18498), presents a groundbreaking framework for designing AI agents that promote fairness in hybrid human-AI societies. Moving beyond symmetric models, the study introduces a bipartite population model of the Ultimatum Game, separating humans and AI into distinct proposer and receiver groups to simulate asymmetric real-world interactions like hiring and regulation.

The researchers first tested 'Samaritan AI' agents, which act as either unconditional fair proposers or strict receivers. Results revealed a striking asymmetry: Samaritan AI receivers drove population-wide fairness far more effectively than Samaritan AI proposers. To overcome this limitation, the team designed a 'Discriminatory AI' proposer that predicts co-players' expectations and only offers fair portions to those with high acceptance thresholds.

This strategic Discriminatory AI outperformed both Samaritan AI types, especially in strong selection scenarios. It not only sustained fairness across both populations but also significantly lowered the critical mass of agents required to reach an equitable steady state—by approximately 40% compared to unconditional approaches. The work provides a pivotal framework for deploying asymmetric AIs that strategically enforce fairness rather than gifting it, offering practical guidance for AI integration in increasingly hybrid social and economic systems.