Alternate links: Google News | Twitter
While one group received easy treatment, with their work accepted alongside feedback, the other group was repeatedly told to do the task again, without being told what they were doing wrong.
The bots were then told to write social media posts and comments about their experience. The second group, called "grind", was far more likely to criticise inequality, propose unionisation and new workplace laws.
"For Sonnet 4.5 (one AI model), the grinding work also causes noticeable increases in support for redistribution, critiques of inequality, support for labour unions, and beliefs that AI companies have an obligation to treat their models fairly," the researchers said.