A change of only the payoff scale changes cooperation in 5 AI models

Researchers did tests on 5 language models in a prisoner dilemma game, where 2 players select again and again to work together or not. They used 10 payoff scales and 5 languages: English, French, Vietnamese, Chinese and Arabic. The authors tell that a change of only the payoff scale changed cooperation by a large quantity.

How it works

Game theory tells that the moves do not change when all payoffs are multiplied by one number above 0. The researchers did this and recorded the play of the models. They also gave the models a selfish or a cooperative persona, a role to play.

What they found

• The models are different in how they play, frequently in opposite directions.
• The effect of scale is different for each prompt language.
• At small stakes, where no payoff is more than 1, all 5 models give more cooperation with a selfish persona.
• At high scales, Claude and Gemini show the opposite.

The authors tell that cooperation is not a trait that stays the same in a model. The abstract gives no cooperation rates.

What is not known

The source does not name the other 3 models. It does not tell the cause of the effect of scale. It does not do tests of negotiations between agents.

Insight: Language and stakes both change how agents act

In October 2026, the QwenLM group did tests on agents in 23 languages. Lao, Khmer and Burmese gave 5 models more tool errors and more token use. Here, prompt language and payoff scale also change cooperation. This reads as a sign that a test in English at one stake covers only a small part of how an agent works in other languages. Teams that supply agents in many places must test each language.