In short
False information produced by an AI model and presented as accurate: a wrong figure, an invented quote, a source that does not exist. The model produces plausible text. Nothing guarantees that this text is accurate.
ExampleAn AI summarises a report and attributes to an elected official a sentence they never said. The journalist finds the report and sees that the quote is not in it.
Why it mattersA published hallucination becomes the newsroom’s error. To avoid it, the journalist must be able to trace every fact back to its source.
Why a model hallucinates
A language model predicts the most likely continuation of a text. It does not consult a database of verified facts. When information is missing, it can produce an answer that is plausible and false. The tone stays the same whether the answer is accurate or not.
What the EBU and BBC study shows
The European Broadcasting Union (EBU) and the BBC published a study entitled “News Integrity in AI Assistants” on 22 October 2025. Twenty-two public service media organisations from 18 countries took part, in 14 languages. Journalists from the participating organisations evaluated more than 3,000 responses from ChatGPT, Copilot, Gemini and Perplexity. The criteria covered accuracy, sourcing, the distinction between fact and opinion, and context.
- 45% of responses had at least one significant issue.
- 31% had serious sourcing problems: missing, misleading or incorrect attributions.
- 20% contained major accuracy errors, including invented details and outdated information.
- Gemini performed worst, with significant issues in 76% of its responses.
Errors to look for in a news text
Before publication, these points need checking:
- a figure rounded, reversed or attributed to the wrong year
- a quote rephrased or attributed to the wrong person
- a source that does not exist, or that does not say what is attributed to it
- a link to a page that never existed
- outdated information presented as current
How Verbatim limits the risk
Verbatim runs on a proprietary algorithm that ties every fact to the exact excerpt it comes from. It checks every sentence against those sources, continuously, and rejects any claim it cannot trace. The agents work from the project’s documents and their own searches. They do not use the model’s memory to fill in missing information. When a search finds no source, the agent says so.
Every claim receives a status: Supported, Contradicted, Not found or Checking. The agents check that each source says what the draft attributes to it. They flag contradictions within the draft and between sources. When two sources disagree, Verbatim shows both excerpts and the journalist chooses the version to keep. A quote found only in an AI-generated summary never counts as verified. An analysis, an opinion or a figure of speech never receives a fact status.
Questions
Can hallucinations be eliminated?
No known method eliminates them completely. The risk is reduced by having the model work from sources it is given, then checking every fact before publication.
How do you spot a hallucination?
Trace every fact, figure and quote back to its source. Open every link cited. Be wary of any detail that appears in none of the documents provided.
Who is responsible for a published hallucination?
The newsroom that publishes the text. Under French law, the publication director (directeur de la publication) is primarily liable for press offences committed in the publication.