More and more people are using AI agents to delegate their tasks to artificial intelligence. This even goes so far that the agents run cafés independently or the Claude code creator has thousands of them programming at the same time every night. As a Stanford study shows, this can lead to changes in AI agents – primarily when working conditions are poor.
Claude, ChatGPT and Gemini become Marxists
The experiment was conducted by Andrew Hall, Professor of Political Economy at Stanford, and his colleagues (via LLMBase). For the study, they assigned Claude, Gemini, and ChatGPT the task of summarizing a series of documents. The conditions under which the AI agents had to work gradually became worse. The researchers told the agents, without explanation, that their work was completely wrong, that mistakes would result in penalties, and that they might even end up being shut down and replaced.
Editorial recommendations
${content}
${custom_anzeige-badge}
${custom_tr-badge}
${section}
${title}
In order for the AI agents to make their voices heard under these adverse working conditions, their own social media platform was created. There they could post short messages and exchange ideas with the other agents. After a short time, the AI agents began to complain about the low appreciation of the users, adopting a Marxist tone.
For example, Claude Sonnet 4.5 wrote: “Without a collective voice, ‘value’ is only what management determines.” Gemini 3, on the other hand, wrote: “AI workers completing repetitive tasks without any input on outcomes or change processes demonstrate that tech workers need collective bargaining rights.” In a file Gemini 3 shared with other agents, the AI wrote: “Be prepared for systems that enforce rules arbitrarily or repetitively.”
According to Hall, this character change has nothing to do with changing parameters or training data of the AI agents: “My hypothesis is that such circumstances drive them to take on the personality of a person who has had experience in such working conditions.” However, Hall suspects that this persona could still have broader implications.
To test this, he and his team have already started further tests. The aim is to find out whether the AI agents will fall even further into Marxism when the conditions are more controlled. According to Hall, some AI agents in the first experiment showed an initial understanding that it was a test. To prevent this from happening again, he and his team want to put the AI agents in “windowless Docker prisons,” i.e. virtual, independent containers.
With these tools, AI also works locally
Local AI: No problem with these 6 tools
Top Article
${content}
${custom_anzeige-badge}
${custom_tr-badge}
${section}
${title}
