One persona change moved GPT-5.2 cooperation by 75.2 points

A team of researchers tested how small changes in an AI social simulation change the result. It used a game of cooperation with turns that go again and again and an echo chamber on social media, with 4 model families. For GPT-5.2, a change to only the representation of a matched persona moved cooperation by 75.2 percentage points.

How it works

A butterfly effect is a small local change that becomes a different result for the full system. The team changed the representation of the persona, the game frame and the memory. Then it compared the results of the group.

The numbers

The frame and the memory gave less effect, or effects that were not the same each time. In a system with many agents, the effect can become more or less, by model. In the echo chamber, the persona changed who agents speak to and the shape of the network.

What is not known

The team gave the 75.2 figure for GPT-5.2 only. It gives TRAILS, a list of robustness checks, and a Social Simulation Audit Card. No other group tested them.

Insight: One simulation can show how it was made, not the world

Teams use AI agents to test how persons act, and the results go into science results. Here one persona change moved most of the cooperation in one game. No one made this effect on purpose. One simulation reads as a sign of how the system was made, more than of how persons act. Tell teams to do it again.