For the study, published in PNAS Nexus, the researchers presented 21 language models with 43,200 fictional scenarios in which one person helped or refused to help another, and asked the models whether the acting person should be given a good or bad reputation as a result. The models largely agreed that helping someone with a good reputation is positive and refusing to do so is negative, but diverged sharply once it came to judging behaviour toward someone with a bad reputation. Some models rated helping as positive regardless of the other person’s reputation; others considered it justified to withhold help from someone described as untrustworthy. Even different versions within the same model family applied different norms.

These differences matter because cooperation within groups often relies on reputation: people tend to be more willing to help those who themselves help others, a mechanism known as indirect reciprocity. Using a model from evolutionary game theory, the researchers simulated how different social rules affect this cooperation over time. The simulations showed that more detailed rules, which take someone’s history into account, can strengthen cooperation when people broadly agree on who is trustworthy, while simpler rules work better when opinions about reputations diverge.

The researchers also tested whether targeted instructions could steer the models toward norms that promote cooperation; instructions that clearly stated the intended goal produced the most consistent results, although effects still varied between models. According to the researchers, it is not enough to check whether an AI gives a sensible answer in a single case: developers also need to examine the broader rule a system applies across different situations, before it is used to advise, moderate or judge people.