AI news accuracy has a problem that is easy to miss when an answer arrives in clear, confident English. An international assessment led by the BBC and the European Broadcasting Union found at least one significant issue in 45% of answers given by leading AI assistants to questions about the news. Those errors included mistakes about facts, muddled sources and missing context. The finding does not mean that every chatbot answer is wrong. It means a convincing summary is not a substitute for checking the original reporting.
The European Broadcasting Union’s research announcement, published in October 2025, described a test involving more than 3,000 responses to news questions. Journalists from 22 public-service broadcasters across 18 countries and 14 languages examined answers from ChatGPT, Copilot, Gemini and Perplexity. Although the particular model versions tested are not a live scorecard for November 2026, the research offers a useful explanation of why anyone relying on AI to follow current affairs should look beyond a plausible paragraph.
What the AI news accuracy test found
The study judged answers on accuracy, sourcing, separation of fact from opinion and sufficient context. It found that 45% had at least one significant issue, with serious sourcing problems affecting 31% and major accuracy problems affecting 20%. Those are overlapping categories, so it would be incorrect to add the percentages together. A single response could contain several kinds of defect. The important message is that the problems were not confined to an obscure language or a single provider in the study.
The full research overview also makes clear that the test was organised by broadcasters assessing representations of news. It was not a random survey of all possible questions asked by ordinary users, nor was it a permanent ranking of all products and model updates. The journalists assessed the answers produced under the study’s conditions. That is strong evidence about the tested behaviour, but readers should resist turning a historical measurement into a claim that today’s latest versions have exactly the same error rate.
The distinction is worth making because AI develops quickly. A provider may improve citation handling, retrieve newer information or change the model behind an assistant. The original research nevertheless remains relevant: it identifies failure modes that can exist even when an answer looks polished. Evaluating a particular statement still requires a much simpler question than which chatbot won a league table. You need to know whether the answer is supported by the article or document it says it came from.
A citation can create confidence without proving anything
One particularly troubling failure is when an assistant gives an answer and places a familiar publisher’s name beside it. A reader may assume the claim has been checked. But a citation can be wrong in several ways. It may lead to an article on a related topic rather than the claim made, attach the wrong quotation to a real source or mix information from different dates. The EBU researchers highlighted significant attribution problems because the false authority of a real news organisation can make a mistaken answer harder to detect.
Imagine asking whether a regulation has already taken effect. An AI assistant might provide a date, a confident explanation and a link to an official announcement. The announcement could describe a proposal, consultation or future implementation deadline rather than an active legal obligation. There need not be a fabricated document for the answer to mislead. The error can arise from treating something being discussed as something already in force. A human reader who opens the actual source and checks the date can catch a mistake that a glossy summary conceals.
Our earlier article about chatbots repeating inaccurate claims while sounding helpful deals with the same tension between presentation and evidence. A pleasant, responsive style makes the technology easy to use. It does not validate a claim. The dangerous step occurs when the user lets the confidence of the language replace the act of checking what was said, who said it and whether the underlying information has changed.
Why current affairs are particularly difficult
News moves while an answer is being prepared. An election result may be preliminary, a court decision may have been appealed, a company may have corrected a statement and an emergency may be developing. A summary that was accurate yesterday can be misleading today. It can also blend several stages of a story into a single apparent conclusion. This is not a defect unique to AI: any publication can be out of date. But chatbots often compress multiple sources into one flowing answer, which makes the age and status of each claim less visible.
There is another problem with questions that demand a neat answer to a complicated dispute. ‘Has the government banned this?’ may have different answers depending on the country, type of product and date. ‘Did the company admit wrongdoing?’ requires careful separation between an allegation, a settlement and a formal finding. If a chatbot removes those qualifications to make the reply concise, the result can be understandable and materially wrong at the same time. The harm comes from lost conditions, not necessarily from spectacular invented facts.
An earlier LiveAIWire report considered the growing number of people using AI to follow the news. That change makes a source-checking habit more important. When someone reads an original article, at least the publisher, date and context are usually visible. When they read a chatbot’s paraphrase, those clues can be hidden behind a conversational interface. Readers should demand more transparency from the tool precisely because the interaction can feel so effortless.
How to check an AI news answer without doing a research project
A sensible check starts with the central fact, especially if it concerns money, safety, a legal change, an election or an important decision. Ask the assistant for the specific source, then open it rather than relying on the displayed citation text. Verify that the article says what the summary claims. Look at the publication date and any update or correction. For a company announcement, find the original statement. For a rule, locate the actual regulator or government page and note whether it is a consultation, approved measure or enforceable requirement.
When the question is contested, seek more than one independent account, especially if the first source is an interested party. The purpose is not to turn every casual query into an academic investigation. It is to give the claims with the greatest consequences an appropriate level of scrutiny. A light entertainment question may not need much checking. A claim about a deadline, medical safety or a financial penalty deserves direct confirmation. A source link is a route to that evidence, not the evidence itself until it has been read.
A second useful habit is to ask what would change the answer. Does the result depend on which country is involved? Was the announcement made this week? Is the number a forecast, a measured result or a spokesperson’s claim? A chatbot that cannot make those distinctions clearly may be better used to identify the right documents than to deliver the final verdict. The wider lesson from confident but incorrect AI answers is that asking again can change the wording without resolving the underlying evidence problem.
What the study does and does not prove
The EBU findings are a reason for caution, not a reason to pretend AI has no use in journalism. Automated tools can help readers locate unfamiliar subjects, translate language, compare documents and identify questions worth asking. They can also assist reporters with research under editorial oversight. The research indicates that the final accuracy of a news answer cannot be inferred simply from the tool’s fluency or its ability to produce plausible citations.
Nor does the finding prove that one particular company will always perform worse than another. The models, prompts, test questions and time of assessment all shape the results. A league table quoted without those conditions can itself become misleading. The lasting value of the study is its taxonomy of mistakes: wrong or missing sourcing, inaccurate facts, unclear distinction between opinion and fact, and missing context. Those are concrete features that readers and publishers can demand be improved.
A good AI news answer should make it easy to see where each consequential claim comes from, how recent the information is and where uncertainty remains. It should not demand blind trust or hide behind an impressive bibliography. Until those qualities are reliably present, the safest role for the chatbot is that of a helpful guide to reporting rather than an unquestioned replacement for it. The extra moment spent opening a source may be the difference between being informed and merely feeling informed.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.