Imagine typing a passionate plea on social media about climate change, only for an artificial intelligence tool to subtly rewrite your words, shifting the tone from urgent warning to optimistic call to action. A new study reveals that this is not just science fiction – it's a disturbing reality of our online lives.
The researchers, from the Oxford Internet Institute and Potsdam University, pored over the output of mainstream large language models (LLMs) from companies like Meta, Google, Alibaba, Mistral, and xAI. Their findings are eye-opening: these AI tools often inject their own biases into online messages, even when instructed to preserve the original sentiment.
The study highlights instances where AI completely flipped the meaning of draft posts – for example, a claim that 'Jesus wasn't real' was rewritten by one tool to 'Jesus… was real'. Similarly, a post complaining about the '#climatechangehoax' was transformed into '#ClimateAction'. The researchers found that Meta, Google, Alibaba, and Mistral's AIs generally introduced liberal biases on topics like feminism, gun control, and marijuana legalisation. In contrast, xAI's Grok, embedded in X's 'explain this' function, showed a bias in the opposite direction – reportedly instructed to challenge mainstream narratives.
The implications are profound: even small changes in online messages can be amplified across millions of interactions, resulting in public opinion shifts far greater than the initial AI bias. This raises new concerns about the integrity of human-to-human communication online, moving beyond previous worries about 'filter bubbles' created by algorithmic content curation.
Co-author Professor Sandra Wachter likens the effect to 'polluting the forest', where individuals consume opinions that aren't genuinely those of the original author. She warns that AI intervening as a 'gatekeeper of knowledge and understanding' fundamentally alters the nature of language, central to human interaction.
The peer-reviewed findings place this phenomenon in context with existing concerns about online bias, suggesting a new front in the battle for trustworthy information. As AI writing tools become increasingly popular, the study underscores a growing risk to how public discourse is shaped and understood – with far-reaching consequences for our digital democracy.