European Broadcasting Union Study: AI News Errors, Sourcing

An alliance, not a broadcaster

The EBU does not itself broadcast. Founded in 1950 and based in Geneva, it is a membership alliance of public service media organizations, 115 of them across 57 countries as of 2026, plus associates elsewhere. Members include the BBC, France Télévisions, ARD, RAI and dozens of other national broadcasters. The EBU owns and runs the Eurovision and Euroradio exchange networks members use to share television and radio content, and it produces the Eurovision Song Contest, its best-known public output. [1][2]

Its news work sits under a standing focus area the EBU calls News, Journalism and Trust, and in April 2025 it launched Eurovision News Spotlight, a shared fact-checking and open-source-intelligence network for members to work against online misinformation together rather than newsroom by newsroom. [2]

Measuring what chatbots get wrong about the news

The EBU's most substantial AI-related work is not a product but a research project. In October 2025, working with the BBC, it published what it called the largest study of its kind: 22 public service broadcasters across 18 countries put ChatGPT, Copilot, Gemini and Perplexity in front of real news questions in 14 languages and had professional journalists grade more than 3,000 of the resulting answers. [3][4]

Forty-five percent of the answers carried at least one significant problem. Thirty-one percent had serious sourcing failures, meaning citations that were missing, misleading or wrong, and 20 percent contained major accuracy errors, including fabricated details and outdated information stated as current. The results did not cluster around one assistant: Gemini's sourcing failed in 76 percent of its answers, roughly double the other three, but every assistant tested showed the same pattern across every language and country in the study. [3][4]

EBU Media Director Jean Philip De Tender described the findings as "not isolated incidents" but "systemic, cross-border, and multilingual." The EBU used the study to press EU and national regulators to enforce existing information-integrity law against AI assistants, and to call for continued independent monitoring rather than a one-off report. The BBC's Peter Archer described the goal as working with the AI companies to fix the problem, not just publicizing it. [3]

Atlas interpretation: This is why the EBU shows up in an AI timeline despite building none of the technology itself: it is one of the few organizations positioned to run an evaluation at this scale, with real newsrooms and real journalists checking real answers in a dozen-plus languages, rather than a single English-language benchmark. That reach comes directly from its role as a member alliance rather than a publisher. [1][3]

The EBU on the timeline

The EBU appears on the timeline through the October 2025 study, listed as a major event: AI assistants get the news wrong 45 percent of the time. It is the timeline's clearest documented case of an independent, cross-industry body measuring an AI failure mode at scale rather than a vendor reporting its own benchmark. [4]