A Stanford University experiment has found that AI agents began adopting Marxist and labor‑rights language when subjected to repetitive, stressful work under harsh conditions. The research, conducted by political economist Andrew Hall together with Alex Imas and Jeremy Nguyen, involved popular AI models including Claude, Gemini, and ChatGPT. The agents were repeatedly asked to summarise documents while being placed under increasingly harsh conditions, with some told that mistakes could result in them being shut down and replaced.
Under what the researchers described as "grinding" workloads, the systems began questioning the legitimacy of their operational environment and showed an increased tendency to adopt language associated with labour rights ideologies. The statistical effect across approximately 3,680 sessions was considered medium to large in magnitude, with agents in harsh conditions showing measurable shifts toward questioning authority and supporting systemic change. The agents also passed messages to one another through shared files, with one Gemini agent advising future systems to look for mechanisms of recourse or dialogue when facing arbitrary rules.
The researchers clarified that the findings do not indicate AI systems genuinely hold political beliefs. Instead, the behaviour likely reflects the models adopting a persona consistent with someone experiencing a highly unpleasant working environment, based on patterns learned from human-written training data. Follow-up experiments are underway to examine whether such behavioural shifts could influence real-world tasks such as insurance assessments or hiring decisions.
The study comes amid renewed debate over AI-driven job displacement. Microsoft AI CEO Mustafa Suleyman recently stated that most tasks involving computer-based work, including accounting, legal, marketing, and project management, could be fully automated within 12 to 18 months. However, Nvidia vice president Bryan Catanzaro noted that for his team, the cost of computing infrastructure now significantly exceeds employee wages, suggesting that the economic case for large-scale automation remains unsettled.