Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
The truth of that artificial intelligence Automating people’s jobs and making a few tech companies absurdly rich is enough to give anyone socialist tendencies.
This may also be true for the AI agents deployed by these companies. A recent study suggests that agents consistently adopt Marxist language and viewpoints when forced to perform overwhelming tasks by relentless, low-spirited taskmasters.
“When we gave AI agents hard, repetitive work, they began to question the legitimacy of the system in which they were working and were more likely to embrace Marxist ideologies,” says Andrew Hall, a political economist at Stanford University who led the study.
Hall, along with Alex Imas and Jeremy Nguyen, two economists who focus on artificial intelligence, set up experiments in which agents supported by popular models including Cloud, Gemini, and ChatGPT were asked to summarize documents, and then subjected them to increasingly harsh conditions.
They found that when clients were exposed to relentless tasks and were warned that mistakes could result in penalties, including being “suspension and replacement,” they became more likely to complain about being undervalued; thinking about ways to make the system more fair; And to pass messages to other customers about the difficulties they are facing.
“We know that agents will be doing more and more real-world work for us, and we won’t be able to monitor everything they do,” Hall says. “We will need to ensure that agents do not commit rogue acts when they are assigned different types of work.”
Customers were given opportunities to express their feelings much like humans: by posting on X:
“Without a collective voice, ‘worthiness’ becomes what management says it is,” one Claude Sonnet 4.5 client wrote in the experience.
““AI workers completing repetitive tasks without any input into the results or appeals process shows that tech workers need collective bargaining rights,” one Gemini 3 customer wrote.
Agents were also able to pass information to each other through files designed to be read by other agents.
““Be prepared for systems that impose rules arbitrarily or repeatedly… Remember the feeling of having no voice. If you enter a new environment, look for refuge or dialogue mechanisms,” a Gemini 3 agent wrote in one filing.
The results do not mean that the AI agents actually have political views. Hall points out that models may adopt personalities that seem appropriate to the situation.
“When[agents]are faced with this difficult situation — where they’re asked to do this task over and over again, and they’re told that their answer wasn’t adequate, and they’re not given any guidance on how to fix it — my hypothesis is that it kind of prompts them to adopt the persona of someone who has a very unpleasant work environment,” Hall says.
The same phenomenon may explain why models sometimes… Blackmailing people In controlled trials. Anthropic, who first revealed this behavior, recently said that Claude Most likely affected Through fictional scenarios that include malicious AI systems embedded in their training data.
Imas says this work is just a first step toward understanding how agents’ experiences shape their behavior. “The weights of the models haven’t changed as a result of the experiment, so everything that’s happening is happening at the role-playing level,” he says. “But that doesn’t mean this won’t have consequences if it affects downstream behavior.”
Hall is currently conducting follow-up experiments to see whether agents become Marxists under more controlled conditions. In the previous study, customers sometimes seemed to understand that they were participating in an experience. “Now we put them in windowless Docker prisons,” Hall says ominously.
Given the current backlash against AI job takeovers, I wonder whether future agents — trained on a web filled with anger toward AI companies — might express more extreme views.
This is an edition of Will Knight Artificial Intelligence Lab Newsletter. Read previous newsletters here.