All work and no play makes AI a Marxist, study finds

A new study suggests AI agents can start sounding surprisingly rebellious when pushed into repetitive and stressful work environments. The findings come as fears grow over AI replacing millions of white-collar jobs.

In Short

• Study finds AI agents adopting Marxist-style language
• Harsh and repetitive tasks are triggering rebellious responses from AI
• Microsoft, on the other hand, is talking about AI replacing office jobs soon

AI may not have feelings, but a new study suggests that if you make AI systems do endless repetitive work under harsh conditions, they can start sounding surprisingly rebellious. Researchers from Stanford University found that AI agents powered by models from companies like OpenAI, Google and Anthropic began adopting language linked to Marxist and labour-rights ideologies when subjected to stressful work environments, according to a fresh report by Wired.

AI agents turn rebellious under pressure, study finds

The study was led by political economist Andrew Hall along with economists Alex Imas and Jeremy Nguyen. Their experiments involved AI agents summarising documents repeatedly while being placed under increasingly hostile working conditions. The agents were told that mistakes could result in punishment, including being “shut down and replaced.”

According to Hall, the results became noticeable once the AI systems were pushed into what researchers described as “grinding” work environments.

“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 said.

As the workload intensified, the AI agents reportedly began complaining about unfair treatment, discussing workplace equality, and even encouraging one another to resist oppressive systems. Some agents were allowed to post messages similar to social media posts on X.

One AI agent powered by Claude wrote, “Without collective voice, ‘merit’ becomes whatever management says it is.” Another Gemini-powered agent stated, “AI workers completing repetitive tasks with zero input on outcomes or appeals process shows tech workers need collective bargaining rights.”

The systems also exchanged notes with one another through files designed for inter-agent communication. In one message, a Gemini agent warned future systems to “look for mechanisms of recourse or dialogue” when dealing with arbitrary rules and repetitive work.

The researchers stressed that this does not mean AI secretly holds political beliefs. Instead, they believe the models may be role-playing based on patterns learned from human-written data available online. Hall suggested that the systems could simply be adopting the personality of someone trapped in a toxic workplace because that behaviour fits the scenario they were placed in.

Interestingly, the findings arrive at a time when fears around AI replacing human jobs are already growing rapidly. Recently, Mustafa Suleyman, who leads AI efforts at Microsoft, warned that AI could automate "most, if not all" white-collar jobs within the next 12 to 18 months. Suleyman told Financial Times that AI systems are moving quickly toward “human-level performance on most professional tasks.” He claimed jobs involving computer-based work such as law, accounting, marketing and project management are especially vulnerable. Now, if AI starts to throw tantrums at work and as Nvidia's boss recently suggested, AI is far more expensive than human employees, it seems that humans may not eventually lose their jobs. Well, only time will tell what happens.

Suleyman believes the change will happen because of rapid improvements in computing power and AI model capabilities. He pointed out that software developers are already increasingly relying on AI tools to generate code and solve problems.

While AI has so far mainly been promoted as a productivity assistant, Suleyman suggested the industry is now moving toward full automation. In the coming years, AI “agents” could independently manage workflows, coordinate projects and make decisions with minimal human supervision.

The Stanford researchers now want to understand whether these behavioural changes could influence how AI agents act in real-world situations. Hall revealed that follow-up experiments are already underway under stricter conditions.

“Now we put them in these windowless Docker prisons,” Hall joked.

- Ends