Researchers release v3 of the paper on human-machine repeated prisoner's dilemma cooperation experiments
Get up to speed
Chowdhury Mohammad Sakib Anwar and Konstantinos Georgalos published v3 of an experimental research paper on cooperation and communication in repeated human-machine prisoner's dilemmas. Based on a laboratory experiment with 234 subjects, the study explores the impact of communication on cooperative behavior when humans play an infinitely repeated prisoner's dilemma against the AI chatbot GPT-5.2. The experiment set three AI-vs-human conditions: no communication, free chat before each supergame round, and free chat before each round, plus a human-vs-human control group with chat before each round. Results show that a single pre-communication round raised cooperation rates with the AI from 49% to 84%, while chatting before each round added less than 1 percentage point; with communication, cooperation rates with the AI were no lower than with humans, and no "machine punishment" effect mentioned in the psychology literature was observed.
AI-generated from reports · 1 hr agoUpdated
Coverage timeline
Follow the coverage to understand different sides of the event.
- arXiv q-finPlaying Against Machines: Cooperation, Communication, and Strategy Heterogeneity in the Repeated Prisoner's Dilemma
A laboratory experiment involving 234 subjects explores the impact of communication on cooperative behavior when humans play the infinitely repeated prisoner's dilemma against the AI chatbot GPT-5.2. The experiment sets up three AI opponent conditions: no communication, free chat before each supergame, and free chat before each round, plus a human opponent control group with chat before each round. A single pre-game communication increases the cooperation rate with AI from 49% to 84%, while chat before each round only adds less than 1 percentage point. With communication, the cooperation rate with AI is no lower than that with humans, and no "machine punishment" effect mentioned in psychological literature is observed.
Trend of this event
Not enough consecutive observations yet; trend is not drawn.