New AI models can distort automated negotiations before anyone chooses to use them.
Researchers studied GLEE, an independent benchmark of 587,000 strategic decisions made by 13 large language models across 1,320 matched bargaining, negotiation and persuasion setups. Treating each new model release as an expansion of the available strategies, they ran more than 50,000 comparisons between releases. In many cases, one agent's payoff rose while its counterpart's fell. The paper calls this the "Poisoned Apple effect": a model nobody actually adopts in equilibrium still redistributes gains and losses and can reshape how a regulator designs the market.
That matters for anything from algorithmic pricing to automated contract negotiation, where AI agents increasingly stand in for people. Regulators tend to judge market fairness by who participates and what they know, not by which model version a vendor happens to ship this quarter. The paper's data suggests up to roughly three in ten of these skewed outcomes trace to models nobody even uses, and that restricting which models can be deployed tends to make the effect worse.
Game theory has always warned that more options can make a game worse for everyone; this is the first evidence that warning shows up at benchmark scale in AI-mediated deals.