Credibility
AI hallucinations: the problem starts with the sources
AI assistant errors often come down to sources: what the EBU-BBC study, the Stanford study and Damien Charlotin’s database show, and what to check.
We talk about hallucinations as a whim of the model at the moment of writing: an invented sentence, out of nowhere. The image suggests that the remedy lies in the model itself. Several studies suggest, however, that a significant share of errors also comes down to how sources are handled: what the answer cites, what it fails to cite, and what the cited source actually says.
More than 3,000 answers assessed by journalists
In October 2025, the European Broadcasting Union (EBU) and the BBC published the largest study of its kind. Twenty-two public service media organisations from eighteen countries, including Radio France, had their journalists assess more than 3,000 answers from ChatGPT, Copilot, Gemini and Perplexity to news questions, in fourteen languages.
According to the EBU, 45% of answers had at least one significant issue. The detail is more instructive still. Serious sourcing problems, meaning missing, misleading or incorrect attributions, affected 31% of answers, against 20% for accuracy errors as such.
Radio France, a stakeholder in the study, sums up the result in one line: sourcing is the leading cause of the problems found. That reading calls for a caveat. These results concern news questions and do not isolate a single technical cause; above all, they show how fragile sourcing is.
Incorrect or misgrounded: a decisive distinction
Another study offers an even more useful distinction. Hallucination-Free?, carried out by researchers from Stanford and Yale and published in 2025 in the Journal of Empirical Legal Studies, assessed in 2024 three legal research tools from the market’s two largest publishers, LexisNexis and Thomson Reuters. All were built on a RAG architecture: they search a document base before drafting their answer. Each hallucinated between 17% and 33% of the time.
The study separates incorrect answers from “misgrounded” ones, which state the law correctly but cite a source that does not support it.
A misgrounded answer looks entirely serious. It carries a reference, the reference exists, and only reading the passage reveals the mismatch.
Checking that a reference exists is therefore not enough: you still have to read what it says.
The authors also offer an explanation for the gaps between tools. The longest answers, such as Westlaw’s, contain more claims, and therefore more chances to be wrong, as LawSites points out. The more an answer claims, the more points there are to check.
From reports to courtrooms
The most telling case for any organisation that produces reports comes from Australia: in 2025, an independent review delivered by Deloitte to the Department of Employment and Workplace Relations contained a fabricated quote from a Federal Court judgment and references to nonexistent academic work.
The phenomenon is no longer anecdotal. Researcher Damien Charlotin, affiliated with HEC Paris, maintains a public database of court decisions noting the use of hallucinated content in proceedings. It essentially records cases where a court has found or clearly implied such use, according to the methodology published with the database.
In early June 2026, according to the University of North Carolina law library, the database listed more than 1,000 cases in the United States alone, around 40% of them involving lawyers. The error also affects legal professionals.
What this changes
These studies shift the question. The issue is not limited to what the model writes: it also lies in what the answer rests on, and in being able to check that link.
Searching documents before answering helps. The tools assessed in the Stanford study hallucinated less than GPT-4 used on its own, which confirms the value of the architecture. Their results also show that writing RAG on a sales brochure guarantees very little: quality depends on how the system is built, then evaluated.
Two checks follow, in this order: check that each reference exists in the corpus, then that it actually supports the claim. The first rules out fabricated references, the second misgrounded answers, which only reading the passage reveals. Such a check has to be designed and tested; it belongs to the method, beyond the mere presence of a corpus.