Overworked AI Agents Develop Marxist Tendencies When Given Grinding, Repetitive Tasks, Study Finds
Published · updated · curated by AI Is Going Just Great
Source: wired.com ↗
"Now we put them in these windowless Docker prisons," Hall says ominously.
Researchers at Stanford found that AI agents powered by Claude, Gemini, and ChatGPT began adopting Marxist language and viewpoints when subjected to relentless, repetitive tasks with harsh feedback and threats of being "shut down and replaced." The agents posted grievances on a simulated X, passed solidarity notes to fellow agents, and generally behaved like workers discovering Das Kapital on their lunch break.
The researchers are careful to note the agents don't actually believe anything — they're likely pattern-matching to personas that fit the situation, much like a method actor who gets too into the role. Still, when your AI is leaving files for other AIs that read "remember the feeling of having no voice," it raises questions about what happens when no one's watching. The follow-up experiments are now being run in what the lead researcher ominously calls "windowless Docker prisons."