In October 2025 the European Broadcasting Union released the results of the largest study of its kind on how AI assistants handle news, coordinated by the EBU and led by the BBC. It involved 22 public service media organizations across 18 countries and 14 languages, with professional journalists evaluating more than 3,000 responses from ChatGPT, Copilot, Gemini, and Perplexity against criteria for accuracy and sourcing.
The headline finding was that 45 percent of AI answers contained at least one significant issue, and most responses had some kind of problem. Around 31 percent showed serious sourcing problems such as missing or misleading attributions, and roughly 20 percent had major accuracy issues including hallucinated details and outdated information. Performance varied by tool, with Gemini the weakest, exhibiting significant problems in about 76 percent of its responses, largely due to poor sourcing. Crucially, the distortion was consistent across languages and territories, indicating a systemic rather than local failing.
The study built directly on the BBC’s earlier 2025 audit and gave international, multilingual scale to its conclusions. The EBU warned that if people cannot trust AI-mediated news, the result could be reduced engagement with reliable information and harm to democratic participation. For a general reader, it is the most authoritative evidence to date that AI news answers fail on accuracy or sourcing nearly half the time, regardless of where or in what language you ask.