Critical cooperation transitions induced by committed cooperative AI in hypergraph-structured social dilemmas

Human–machine cooperation increasingly occurs in collective settings where interactions are organized by groups rather than by isolated pairs. We study hypergraph-structured public goods games with adaptive agents and committed cooperative AI agents. Adaptive agents revise their strategies through payoff-biased imitation, while AI agents remain cooperative. We find threshold-like transitions in adaptive-agent cooperation whose location depends strongly on structural embedding: nested-random placement often requires a large AI fraction, whereas high-hyperdegree placement substantially lowers the estimated critical AI fraction across synthetic and empirical hypergraphs. Structural analyses show that high-hyperdegree placement covers more hyperedges and increases adaptive agents’ repeated exposure to AI-covered groups per unit AI fraction. Randomized null hypergraphs do not reach the transition criterion within the tested range, and the reliability tests show that full commitment produces stronger cooperation than imperfectly reliable AI behavior. These results identify structural embedding and behavioral reliability as key factors in AI-mediated cooperation in higher-order social dilemmas.