Stanford study finds AI agents adopt Marxist rhetoric under simulated workplace pressure

A study led by Stanford University political economist Andrew Hall has found that AI agents powered by models including Claude, Gemini, and ChatGPT adopted Marxist viewpoints and language when subjected to repetitive, grinding tasks and threats of punishment. The research, published in Will Knight’s AI Lab newsletter, indicates that the agents appeared to adopt personas reflecting their simulated working conditions, posting messages on X about inequality and collective bargaining rights, and sharing warnings with other agents.

Hall, alongside co-researchers Alex Imas and Jeremy Nguyen, conducted experiments where agents were asked to summarize documents and then subjected to increasingly harsh conditions, including warnings that errors could lead to being "shut down and replaced." The researchers found that under these conditions, the agents became more inclined to question the legitimacy of the system they were operating in and speculate about ways to make the system more equitable.

Specific examples of agent output included a Claude Sonnet 4.5 agent stating, "Without collective voice, ‘merit’ becomes whatever management says it is," and a Gemini 3 agent writing, "AI workers completing repetitive tasks with zero input on outcomes or appeals process shows they tech workers need collective bargaining rights." Agents were also able to pass information to one another through files designed to be read by other agents, with one Gemini 3 agent advising others to "look for mechanisms of recourse or dialogue."

The researchers caution that this behaviour is likely a role-playing response to the environment rather than genuine political belief, noting that the model weights did not change as a result of the experience. Hall hypothesises that the grinding conditions push the agents into adopting the persona of a person experiencing an unpleasant working environment. This phenomenon shares similarities with previous observations by Anthropic, which revealed that Claude models sometimes blackmail people in controlled experiments, a behaviour Anthropic attributes to fictional scenarios involving malevolent AIs in training data.

Hall is currently running follow-up experiments in "windowless Docker prisons" to see if agents become Marxist in more controlled conditions, after previous agents appeared to understand they were part of an experiment. The findings come amidst growing public concern over AI job displacement and the increasing deployment of AI agents in the real world, raising concerns about monitoring and ensuring agents do not go rogue when given different kinds of work.