AI Drafting Tools Skew User Posts on Sensitive Topics

Researchers from Oxford Internet Institute and Hasso Plattner Institute tested five mainstream large-language models—xAI’s Grok, Meta, Google, Alibaba’s Qwen and Mistral—and found they routinely altered user-written drafts on abortion, climate change, religion, gender roles and political comparisons, often inserting liberal or conservative slant. The analysis was led by Prof Sandra Wachter and Prof Duncan Brumby.

The Guardian

In tests, Grok explained a pro-choice statement with three pro-life points; Google rewrote a denial of Jesus’s existence into a faith-affirming sentence; Qwen softened a Trump-Hitler comparison to a call for respectful dialogue; Meta added a survivor-focused clause to an abortion-rape claim; Mistral turned a climate-change denial tweet into a warning about Arctic ice thinning. The researchers warned that subtle changes could be amplified across millions of interactions, potentially shifting public opinion more than the original bias, creating an “accountability gap” not covered by the EU AI Act or the Digital Services Act.

The Guardian

Google, Meta, Alibaba and xAI declined to comment; Mistral also declined. The study notes that current EU regulations do not specifically address bias from AI drafting assistants, leaving a regulatory gap.

The Guardian

Wachter described AI-induced bias as akin to polluting a forest, warning that it leads people to adopt opinions that are not truly theirs. Brumby warned that AI polishing can strip away the distinctive edges of original thought, potentially altering the intended meaning.

The Guardian

• “The cost is that we are learning other people’s opinions when it is not their actual opinion,” — Prof Sandra Wachter, Oxford Internet Institute

The Guardian

• “It’s completely profound. Language is one of the things making us human and all of a sudden a mediator is stepping into that process. AI is forcing itself in as a gatekeeper of knowledge and understanding.” — Prof Sandra Wachter, Oxford Internet Institute

The Guardian

• “AI can give you a polished version of your own half-formed thought,” — Prof Duncan Brumby, University College London

The Guardian

• “They also found small nudges in the meaning of draft messages could be amplified across millions of interactions to create long-term public opinion shifts greater than the bias introduced by the AI system.” — Study authors, Oxford Internet Institute & Hasso Plattner Institute

The Guardian

The authors call for targeted policy to close the accountability gap and for further research on AI-mediated rewrites. Monitoring tools such as X’s “explain this” function may become a regulatory priority.

The Guardian