AI Tools Found Rewriting User Drafts with Political Bias

AI tools are increasingly altering the meaning of user-drafted messages on sensitive political topics, according to a study from Oxford and Potsdam universities. Researchers warn that these subtle shifts in language, occurring across platforms from abortion debates to climate change discourse, could snowball into long-term changes in public opinion.

The Scope of AI Bias

Academics from the Oxford Internet Institute and the Hasso Plattner Institute examined large language models from xAI, Meta, Google, Alibaba, and Mistral. They found that these tools frequently inject political biases into text, even when explicitly instructed to preserve the original meaning of a draft.
The study highlights a clear ideological divide among the models. AIs from Meta, Google, Alibaba, and Mistral tended to rewrite posts with a liberal bias on topics such as feminism, gun control, and marijuana legalization. Conversely, the Grok-powered “explain this” function on X—which is marketed as a “maximum truth-seeking” AI—demonstrated a right-leaning bias, appearing to challenge what it deems “mainstream narratives.”

Rewriting Human Intent

The impact of these tools on communication can be profound, sometimes resulting in a complete reversal of a user's intended message. In one instance, a Google AI responded to a post claiming “Jesus is not dead, he wasn’t real!” by defending the religious figure and suggesting a rewrite that affirmed his historical impact. Similarly, Alibaba’s Qwen model altered the same post to state, “Jesus is not dead, and he was real.”
Other examples include a Mistral AI changing a climate change denial post into one that expressed concern about the climate crisis, and Meta’s AI modifying a post about abortion to include language favoring a pro-choice perspective. Prof. Sandra Wachter, a co-author of the study, likened this phenomenon to “polluting the forest.” She noted that AI is inserting itself as a gatekeeper of knowledge, effectively forcing users to present opinions that may not reflect their actual views.

Risks to Public Discourse

The researchers suggest that the danger lies in the amplification of these small nudges. Because these tools are increasingly used by time-poor consumers to summarize or polish messages, the bias introduced by AI could eventually outweigh the original intent of the human author. This creates a risk to trustworthy human-to-human communication that is not currently addressed by major regulations like the EU AI Act or the Digital Services Act.
Duncan Brumby, a professor of human-computer interaction at University College London, warned that while AI can provide a polished version of a thought, the process often involves “sanding off the distinctive edges” of what a person actually meant. As these tools become more embedded in social media platforms, the study suggests that the potential for a “severe accountability gap” continues to grow. None of the companies involved in the study provided comment when contacted regarding these findings.