Testing for completions that simulate altruism in early language models

Altruism underlies cooperative behaviours that facilitate social complexity. In late 2022 and early 2023, we tested whether particular large language models—then in widespread use—generated completions that simulated altruism when prompted with text inputs similar to those used in ‘dictator game’ experiments measuring human altruism. Here we report that one model in our initial study set—OpenAI’s text-davinci-003—consistently generated completions that simulated payoff maximization in a non-social decision task yet simulated altruism in dictator games. Comparable completions appeared when we replicated our experiments, altered prompt phrasing, varied model parameters, altered currencies described in the prompt and studied a subsequent model, GPT-4. Furthermore, application of explainable artificial intelligence techniques showed that results changed little when instructing the system to ignore past research on the dictator or ultimatum games but changed noticeably when instructing the system to focus on the needs of particular participants in a simulated social encounter.