BBC study finds AI assistants distort news in over half of responses

The BBC published research testing four AI assistants — OpenAI’s ChatGPT, Google’s Gemini, Microsoft’s Copilot and Perplexity — by asking each to summarise and answer questions on 100 BBC news stories using the assistants’ own news-retrieval and citation features. Journalists then reviewed the responses against the source articles. The BBC found that 51% of all AI-generated answers had significant issues of some kind, 19% of responses that cited BBC content contained factual errors — wrong numbers, dates or statements — and 13% of quotes attributed to BBC articles had been altered from the original or did not appear in the cited piece at all.

Error rates varied sharply by assistant: the BBC found problems in 34% of Gemini’s responses that used its content, against 27% for Copilot, 17% for Perplexity and 15% for ChatGPT. Beyond factual slips, the report described assistants that struggled to distinguish opinion from reported fact, added editorialising commentary not present in the source, and omitted context needed to understand a story correctly — examples included chatbots stating that Rishi Sunak and Nicola Sturgeon still held political offices they had left, and Gemini misstating NHS guidance on vaping as a smoking-cessation aid.

Deborah Turness, CEO of BBC News and Current Affairs, wrote that AI companies were “playing with fire” by shipping products that could not reliably represent journalism they were drawing on. The study became one of the most cited pieces of evidence in a growing argument over AI search and chat products acting as intermediaries between publishers and readers, feeding directly into the wave of publisher lawsuits and licensing disputes that continued through 2025.