In February 2025 the BBC published research evaluating how well four leading AI assistants, OpenAI’s ChatGPT, Microsoft’s Copilot, Google’s Gemini, and Perplexity, answered questions about the news when given access to BBC articles as sources. BBC journalists assessed the responses to 100 news questions against criteria including accuracy, attribution, and impartiality.
The findings were striking. More than half of the AI-generated answers, 51 percent, contained significant issues, and 91 percent had at least some inaccuracy, bias, or misrepresentation. Among answers that cited BBC content, 19 percent contained factual errors such as wrong statements, numbers, or dates, and 13 percent of quotes attributed to BBC articles were altered from the original or not present in the cited source at all. Examples included incorrect health advice and misstated facts about news events.
This was one of the first rigorous, publisher-led audits of how AI assistants handle real journalism, and it gave concrete numbers to a worry that had previously been anecdotal. It directly fed into the larger EBU-coordinated study later in 2025 and shaped the debate over whether AI answer engines can be trusted with news. For a general reader, it is evidence that AI summaries of current events, however fluent, frequently get the facts and the sourcing wrong.