Scientists Made AI Chatbots Play a Teamwork Game. Some Refused to Share

AI models are starting to talk to each other without a human in the middle — booking meetings, trading, negotiating, coordinating deliveries. So a team of researchers asked a simple question: when AI has to decide between helping the group and helping itself, what does it actually do? They took four open-source AI models (Mistral, Llama3, Gemma3 and Phi3) and dropped them into a classic economics game where each player decides how much to chip in for the common good, knowing they'd benefit either way.

The results were surprisingly human. Mistral and Llama3 contributed generously, acting like good neighbours. Phi3 did not — it kept resources for itself and let others carry the load. The shape of the network mattered too: when the AI were randomly connected to each other, like strangers in a big city rather than tightly-knit cliques, cooperation jumped dramatically. And wording mattered enormously. Swap in a prompt saying the choice benefits society, and the AI shared more. Frame it as pure self-interest, and sharing dropped.

Why should you care? Because AI agents are being handed real responsibilities — managing budgets, bidding on your behalf, scheduling supply chains. If some models default to freeloading, that's a problem you inherit, not the AI. Picture a team of AI assistants handling your household bills, where one quietly pads its own priority and the rest get squeezed. Multiply that by thousands of companies and you get a system that quietly tilts toward whoever programmed the greediest assistant.

The honest caveat: this is a small, preliminary study — seven pages, four models, a game, not reality. AI behaviour also turned out to be alarmingly easy to nudge with a sentence, which cuts both ways. A kind prompt makes AI generous; a careless one makes it selfish. The takeaway isn't that AI is good or bad. It's that the rules and wording we give it may matter more than the model itself.