Physics of Agents paper: AI groups drift right on politics

A new arxiv preprint simulates 'over 10,000 communities of language-model agents' that exchange messages and revise opinions. On objective questions, communication helped accuracy. On subjective political questions, the groups 'drifted group opinions toward the right.' Two experts in our Who's Who directory have shared the paper.

The authors identify three behavioral regimes (indifference, polarization, consensus) and fit a statistical-mechanics model to the resulting agent trajectories. Their framework, they write, 'outperforms all standard baselines' and generalizes to community structures it was not trained on.

Three fitted parameters carry most of the story. Communities operate 'below the critical social temperature, which explains conviction buildup.' Attractive social ties outweigh repulsive ones, tilting groups toward consensus. Agents holding correct answers exert what the paper calls 'the strongest pull, which drives truth-seeking', the mechanism the authors credit for the accuracy gains on factual questions.

The abstract does not name which language models power the agents, or the size of the rightward political shift. The paper itself runs 51 pages with 20 figures and 9 tables. Its bottom line, in the authors' own words: 'collective behavior of AI agents...follows compact and predictive dynamical laws.'