Nearly half of AI answers about news had significant errors

When you ask an AI assistant about today's news, you can't count on the answer being right. That's the finding of a study by the European Broadcasting Union (EBU) and the BBC, published on October 21, 2025.

Journalists from 22 public service media organizations in 18 countries and 14 languages assessed more than 3,000 responses from ChatGPT, Copilot, Gemini and Perplexity to news questions. They looked at whether the answers were accurate, whether they cited sources, whether they distinguished between opinion and fact, and whether they provided the necessary context.

Every other answer had problems

According to Journalism.co.uk, which reviewed the report, 45 percent of the answers had at least one significant problem. About 31 percent had serious sourcing problems, meaning it was unclear or wrong where the information came from. About 20 percent had major inaccuracies, for example fabricated or outdated information.

Gemini performed worst. 76 percent of its answers had significant problems, mainly because the source citations were poor. The EBU site shows no breakdown for the other assistants.

The EBU concludes that AI assistants routinely misrepresent news content, regardless of language, territory or platform. According to the EBU, this points to a systemic problem rather than isolated errors.

Young people use it most

According to the report, 7 percent of all online news users use AI assistants for news. Among those under 25, the share is 15 percent.

The EBU and the BBC are calling for better AI answers, more media literacy and stricter enforcement of rules on information integrity. They have also published a toolkit that can be used to assess the news integrity of AI assistants.

Sourcing is the weak spot

Tech-insider.org describes another test from the Tow Center that included 200 articles. In it, Perplexity cited the wrong source in 37 percent of cases and ChatGPT Search in 67 percent. Grok 3 came in at 94 percent. The figures come from tech-insider.org's coverage and measure source attribution, not general accuracy.

Tech-insider.org also writes that other benchmarks show roughly the same overall factual reliability for Perplexity, ChatGPT and Gemini. The difference lies mostly in how correctly they cite sources.

The study tested answers from 2025, and the models have been updated since. The results therefore don't necessarily say anything about the current versions.

For you as a reader, this means in practice that you should check the source if an AI answer about news is important to you.