News Integrity in AI Assistants [pdf] (www.ebu.ch)

An international study led by the BBC and the European Broadcasting Union evaluated how four widely used AI assistants — ChatGPT, Microsoft Copilot, Google Gemini and Perplexity — answer news questions. Twenty-two public service media organisations across 18 countries and 14 languages asked 30 core (and additional local) questions and had journalists rate free/consumer-version responses (generated May–June 2025) on five criteria: accuracy (including direct quotes), sourcing, opinion vs fact, editorialisation and context. While there has been some improvement since the BBC’s earlier round (the BBC-to-BBC share of responses with significant issues fell from 51% to 37%), the multi-market audit found errors remain widespread: 45% of responses contained at least one significant issue. Sourcing was the largest problem (31% of significant issues), with Gemini especially poor on attribution (72% of its responses had major sourcing errors) while Copilot, ChatGPT and Perplexity each had roughly a third of responses with significant faults. Technically, the study shows problems are systemic across languages and markets and include misattribution, overconfidence, editorialisation and insufficient context — all of which can mislead users and damage publisher trust as AI “answer-first” experiences divert traffic (publishers report search referral drops of 25–30%). The report releases a “News Integrity in AI Assistants Toolkit” and urges developers to prioritise accuracy and transparent reporting by language/market, give publishers control and clear citation formats, and for policymakers to consider accountability measures. It also calls for continued media literacy work so audiences understand assistants’ limits.