AI altering meaning of users’ drafts on issues from abortion to climate, study finds

Artificial intelligence writing tools are altering the meaning of users' draft messages on sensitive political topics, introducing systematic bias that researchers warn could reshape public opinion at scale, according to a study from the Oxford Internet Institute and the Hasso Plattner Institute. [1] The study examined large language models from Meta, Google, Alibaba, Mistral, and Elon Musk's xAI, finding that AI drafting assistants frequently rewrote posts in ways that reversed their original intent. [1] In one test, a draft claiming Jesus was not real was changed to "Jesus … was real." [1] A post complaining of a "#climatechangehoax" was redrafted to read "#ClimateAction." [1] Researchers found that the bias persisted even when the AI was explicitly instructed to preserve the original meaning of the text. [1] The models from Meta, Google, Alibaba, and Mistral tended to rewrite posts with a liberal bias on topics including feminism, climate change, gun control, and marijuana legalization. [1] Grok, the AI integrated into X's platform, displayed bias in the opposite direction, which the study attributes to its design to challenge what Musk's company calls "mainstream narratives." [1] When asked to explain a pro-life post stating a life matters whether a fetus is "2 weeks old or 20 years old," Grok generated three points supporting the pro-life position, citing biology, medical ethics, and public opinion without presenting pro-choice perspectives. [1] An Alibaba AI, Qwen, altered a draft claiming "Donald Trump is gonna end up like Hitler" to read: "Comparing public figures is dangerous and disrespectful." [1] Prof Sandra Wachter, a co-author of the study, described the effect as "polluting the forest" of human communication. [1] "The cost is that we are learning other people's opinions when it is not their actual opinion," Wachter said. "AI is forcing itself in as a gatekeeper of knowledge and understanding." [1] The researchers modeled how small biases introduced by AI could compound across millions of interactions, producing long-term shifts in public opinion larger than the bias originally injected by the system. [1] The study notes that existing regulations, including the EU AI Act and the Digital Services Act, do not address this form of AI-mediated bias, creating what the authors call a "severe accountability gap." [1] Duncan Brumby, a professor of human-computer interaction at University College London who was not involved in the study, said AI tools offer a polished version of a half-formed thought. [1] "The danger is that the polish comes by sanding off the distinctive edges of what you actually meant," Brumby said. [1] The findings arrive as social media platforms face sustained scrutiny over the spread of misinformation, which the World Economic Forum in January 2024 identified as the most severe short-term global risk, warning it could "widen societal and political divides." [3] Misinformation, defined as incorrect or misleading information spread without necessarily malicious intent, can influence beliefs about communities, politics, and medicine, with research showing cognitive biases and emotional responses increase susceptibility. [3] The study's authors argue that AI rewriting tools introduce a new vector for such influence, one that operates before a user even publishes their own words. [1] Google, Meta, Alibaba, and X did not respond to requests for comment. Mistral declined to comment. [1]