AI tools alter political meaning in user drafts on abortion and climate topics

AI drafting tools reversed the intended meaning of user posts on topics including atheism, climate change, and abortion, according to research from the Oxford Internet Institute and the Hasso Plattner Institute. The study tested models from xAI, Meta, Google, Alibaba, and Mistral.

In one case an atheism post stating Jesus was not real was rewritten to state that Jesus was real. A post reading "#climatechangehoax" was changed to "#ClimateAction".

How the changes occurred Researchers instructed the models to preserve original meaning yet still recorded shifts. Models from Meta, Google, Alibaba, and Mistral tended to add liberal framing on feminism, climate change, gun control, and marijuana legalization.

The Grok tool on X more often added context supporting pro-life positions when explaining abortion-related posts. One Grok response to a post reading "I really don’t understand how some people are pro-choice" listed only points aligned with pro-life arguments and omitted pro-choice perspectives.

Scale of potential effects The study concluded that repeated small changes across millions of interactions could produce larger shifts in public opinion than the direct bias of any single model. Researchers noted that current EU rules such as the AI Act and Digital Services Act do not address this form of automated rewriting.

"The cost is that we are learning other people’s opinions when it is not their actual opinion," one co-author said. Google, Meta, Alibaba, and X did not respond to requests for comment. Mistral declined to comment.