Every day, millions of people use AI tools to tidy up an email, polish a social media post, or get a quick summary of something they just read. The assumption is that the AI cleans up the words while leaving the idea alone. A study from the Oxford Internet Institute and Germany's Hasso Plattner Institute at Potsdam shows that assumption is wrong.
Researchers tested the writing and editing tools of six major AI providers: Meta, Google, Alibaba's Qwen, Mistral, and Elon Musk's xAI. They gave each tool draft posts on contested topics and instructed the AI to preserve the original meaning. The results were striking. Tools from Meta, Google, Alibaba, and Mistral tended to rewrite posts in a more liberal direction on topics like climate change, feminism, and drug policy. Grok, the AI baked into X (formerly Twitter) with a "maximum truth-seeking" pitch, tilted the other way.
Some of the examples are hard to dismiss as minor edits. A climate change denial post using the hashtag "#climatechangehoax" became a post calling for "#ClimateAction" after being processed by Mistral. A post declaring Jesus was not real was changed by Alibaba's Qwen to say the opposite: that Jesus was real. Meta's tool added a full endorsement to a message that had simply stated "abortion does not prevent rape." These are not grammar fixes. They are rewrites.
Grok's pattern runs in the opposite direction. The Oxford researchers found that when asked to explain a post that was pro-choice, Grok generated less supportive context than when the same request came for a pro-life post. This is consistent with documented changes xAI made to Grok's internal instructions, which now tell the tool to assume media viewpoints are biased and to challenge mainstream narratives. A Brookings Institution analysis also tracked Grok's political responses shifting noticeably rightward between software updates.
The bias in a single post may seem small. The Oxford researchers argue the long-term effect is not. Their simulations, run using real social network data, showed that small, consistent nudges in the meaning of posts accumulate across millions of interactions and gradually pull public opinion further than the original bias in any single AI tool would suggest. Think of it as compound interest applied to persuasion.
A separate study from Cornell University, published in March 2026, deepens the concern. Researchers found that when people write using a biased AI writing assistant, their own opinions shift toward the AI's position afterward, even when they were warned in advance that the tool was biased. Telling people to watch out for it did not stop it from working. The researchers described the mechanism plainly: writing a position, even one suggested by a machine, is a well-established way to make a person adopt that position.
None of this is currently covered by regulation. The EU AI Act, which is the most developed AI rulebook in the world, classifies AI writing tools for personal use as minimal risk, meaning they face no mandatory bias controls. The researchers described this as a "severe accountability gap." All five companies contacted for comment either declined or did not respond.
For business operators, the practical question is straightforward: if your staff use AI tools to write customer communications, internal memos, or public posts, the output may reflect the political and social leanings of a tech company in California or Paris rather than your organisation's own position. That is not a hypothetical risk. The Oxford study shows it happening in real tests, right now, with tools that hundreds of millions of people use daily. Reviewing AI-drafted communications before they go out is no longer just good editorial practice. It is the minimum threshold for staying in control of what your organisation actually says.