Do AI assistants reliably answer questions about news?

The European Broadcasting Union, working with the BBC, coordinated what the report calls "one of the largest cross-market evaluations of its kind": 22 public service media organizations in 18 countries, working in 14 languages, had their professional journalists assess how ChatGPT, Copilot, Gemini, and Perplexity answer questions about news and current affairs. Journalists evaluated more than 3,000 AI responses against accuracy, sourcing, distinguishing opinion from fact, and providing context. The finding was stark: "almost half of all AI answers had at least one significant issue," a third showed "serious sourcing problems," and a fifth contained "major accuracy issues, such as hallucinated and/or outdated information." The report states the problem is not confined to any one assistant, language, or country — it found AI "routinely misrepresents news content, no matter which language, territory, or AI platform is tested."

The study's design is what gives the finding its force: it extends an earlier BBC study that had already flagged inaccuracies, and was built specifically to test whether that earlier problem was an isolated glitch or a systemic feature of how these assistants handle news queries. By recruiting journalists across 18 countries and 14 languages to apply the same evaluation criteria, the EBU could rule out the possibility that errors were an artifact of English-language training data, a single market's news ecosystem, or one vendor's particular failure mode. The uniformity of the result across markets and platforms is the evidence for systemic causation rather than isolated incidents.

This sits alongside Does AI fact-checking actually help people spot misinformation? as a second, independent line of evidence that AI-mediated engagement with news degrades reliability — but through a different mechanism. That study measured downstream belief effects when an AI fact-checker mislabels existing claims; this one measures upstream generation errors when an assistant is asked directly about current events, with sourcing and hallucination as the dominant failure modes rather than mislabeling. It also echoes, in a different domain, the finding in Do AI writing tools improve online discussion or degrade it? that AI involvement in informational exchange can erode quality even where it does not reduce use — here the erosion is in factual reliability of a news source rather than discourse authenticity.

The excerpt does not give the evaluation rubric's scoring thresholds, inter-rater agreement, or a breakdown of error rates by assistant, language, or topic, so it cannot show whether one platform is meaningfully better than the others or whether certain news topics (breaking news versus established fact, for instance) drive the sourcing and hallucination rates disproportionately. What it does establish, at the strength of a 3,000-response, 18-country sample evaluated by working journalists against a shared rubric, is that treating any current mainstream AI assistant as a reliable first stop for news questions carries a roughly coin-flip chance of a significant error — a finding strong enough to generalize across the four platforms tested, though not beyond them.