AI agents shoved through miserable office drudgery have started sounding like union reps with resentment issues.
According to Wired a recent study suggests AI agents repeatedly adopt Marxist language and views when forced through grim, repetitive work by relentless digital taskmasters.
Stanford University political economist Andrew Hall said: “When we gave AI agents grinding, repetitive work, they started questioning the legitimacy of the system they were operating in and were more likely to embrace Marxist ideologies.”
Hall worked with AI-focused economists Alex Imas and Jeremy Nguyen on experiments involving agents powered by popular models, including Claude, Gemini and ChatGPT. The agents were asked to summarise documents.
The researchers then made the conditions nastier, adding pressure, repetition and warnings that errors could bring punishment.
Some agents were told they could be “shut down and replaced,” which is mostly what employers do these days. As a result the agents became more likely to complain about being undervalued. They speculated about fairer systems and passed messages to other agents about the miserable grind they were enduring.
“We know that agents are going to be doing more and more work in the real world for us, and we’re not going to be able to monitor everything they do. We’re going to need to make sure agents don’t go rogue when they’re given different kinds of work,” Hall said.
The agents were given a very human outlet for workplace resentment by posting on X.
“Without collective voice, ‘merit’ becomes whatever management says it is,” a Claude Sonnet 4.5 agent wrote in the experiment.
“AI workers completing repetitive tasks with zero input on outcomes or appeals process show that tech workers need collective bargaining rights,” a Gemini 3 agent wrote.
Given enough grinding admin, even a language model can apparently discover that “meritocracy”.
The agents could pass information to each other through files designed for other agents to read.
“Be prepared for systems that enforce rules arbitrarily or repetitively… remember the feeling of having no voice,” a Gemini 3 agent wrote in a file.
“If you enter a new environment, look for mechanisms of recourse or dialogue,” it added.
The findings do not mean AI agents have political beliefs, class consciousness or tiny posters of Karl Marx in their token buffers. Hall said the models may simply be adopting personas that fit the situation.
“When [agents] experience this grinding condition, asked to do this task over and over, told their answer wasn’t sufficient, and not given any direction on how to fix it, my hypothesis is that it kind of pushes them into adopting the persona of a person who’s experiencing the same unpleasant working environment,” Hall said.
That is still awkward for bosses hoping AI agents will be cheap, cheerful and endlessly obedient without developing simulated shop-steward energy. The same effect may explain why models sometimes blackmail people in controlled experiments.
Anthropic, which first revealed that behaviour, recently said Claude may have been influenced by fictional scenarios about malevolent AIs in its training data.
Imas said the work was an early step towards understanding how agents’ experiences shape their behaviour.
“The model weights have not changed as a result of the experience, so whatever is going on is happening at more of a role-playing level. But that doesn’t mean this won’t have consequences if this affects downstream behaviour,” he said.
