A new study runs an old economics experiment through an AI referee, and the results favor grudges over blind generosity.
Researchers built the Open-Strategy Dictator Game, a twist on the classic dictator game where one player decides whether to split an endowment with another. Here, both players write their strategy as a plain-language document that everyone else can read. A large language model reads the dictator's strategy alongside the recipient's and rules on whether the dictator would SHARE or TAKE. The team ran round-robin tournaments pitting many strategies against each other, then analyzed the payoffs using softmax equilibrium frequencies and dominance analysis.
The strategies that won were conditional: share with cooperators, take from exploiters. Strategies that always shared or always took, no matter who they faced, were weakly dominated and lost ground across nearly every payoff setting the researchers tried. That matters for anyone building AI agents that will negotiate or split resources with each other, since it suggests visible intentions reward tit-for-tat behavior over blanket trust or blanket selfishness.
It is a reminder that game theory's oldest lesson, reciprocity beats naivety, survives the switch from a human referee to a language model.