Study shows AI assistants misrepresent news content 45% of the time (www.ebu.ch)

An international study coordinated by the European Broadcasting Union and led by the BBC tested over 3,000 AI assistant responses across four major systems (ChatGPT, Copilot, Gemini, Perplexity) in 14 languages with contributions from 22 public-service media organisations. Professional journalists assessed answers against accuracy, sourcing, separation of opinion and fact, and contextualising news. Overall, 45% of responses contained at least one significant issue: 31% had serious sourcing problems (missing, misleading, or incorrect attribution), 20% had major accuracy faults (hallucinated or outdated details), and Google’s Gemini performed worst with significant issues in 76% of its answers. The study found improvements versus earlier BBC research but concluded problems remain widespread and multilingual. This matters because AI assistants are increasingly replacing search as a news gateway—7% of online news consumers and 15% of under-25s use them—so systemic distortions risk eroding public trust and democratic participation. The EBU/BBC team released a “News Integrity in AI Assistants Toolkit” to define what good responses look like and propose fixes, and are urging regulators to enforce existing rules while pursuing ongoing independent monitoring. The findings underscore that technical work on provenance, real-time sourcing, model grounding, and better evaluation frameworks is urgent if AI assistants are to be reliable news intermediaries.