Fact-Checking AI Output Before You Publish

The error that is no longer rare

There is a public database collecting court cases in which the court itself established that a filing contained AI-invented citations. By early September 2026 the number of entries was above two thousand, from more than forty jurisdictions. According to the person who maintains it, before 2025 such cases arrived at two or three a month; by February 2026 it was around five a day. And that is an undercount: only cases a court has already ruled on go in.

The procedural consequences have arrived too. In an appellate decision of March 2026, two lawyers were fined $15,000 each because their brief contained "over two dozen fake citations and misrepresentations of fact" — which the court called a conservative estimate. One sentence from the reasoning is the whole point: no filing should contain any citation that counsel has not personally read and verified. The courts of one US state have gone as far as a per-item schedule: $500 per fabricated citation, $1,000 per fabricated quotation.

Science shows the same curve. An analysis published in May 2026 verified 97 million references by identifier: in 2023 one paper in 2,828 contained a fabricated reference, in 2025 one in 458, and in the first weeks of 2026 one in 277. A twelve-fold rise in two years — and a floor rather than a ceiling, because references without a resolvable identifier were excluded from the study.

Why asking the model does not work

In the first case that drew public attention, back in 2023, the lawyer asked the chatbot whether the cited cases were real. The answer was yes, and that they could be found in the major legal databases. None of them existed. This is not an anecdote but a perfect illustration of the failure type: self-checking comes from the same system as the error.

Measurement says the same. In a controlled study, self-correction without outside feedback degraded performance on every benchmark: on one mathematical task from 95.5 percent to 89 over two rounds, on another from 75.8 to 38.1. A model fails to correct itself not out of laziness but because it has no external yardstick to correct against.

Error recognition is selective too, and weak in exactly the wrong place. In a 2025 measurement, models recognised 58 percent of their own hallucinations when asked to work step by step — but the hit rate was above 80 percent for geographic names and companies and below 40 percent for people and entertainment. In other words, it is weakest precisely where invention is most likely. Why all this is so is covered in hallucination explained — briefly: evaluation has traditionally rewarded guessing over saying "I don't know".

What changed: the fake citation is no longer the main problem

This is the part that is different in 2026 from two years ago. A study published in July 2026 requested 2,736 citations from a current model against clinical guidelines: 7.13 percent were outright fabricated — a striking improvement on the 18 to 55 percent of earlier models. But in that same sample only 49.34 percent were bibliographically accurate, and only 8 percent of queries returned a flawless citation list. The typical defects: the wrong identifier in 37 percent of cases, the wrong title in 31, the wrong page numbers in 27.

Another measurement, from May 2026, shows this even more sharply on research agents. With frontier models the cited link worked in 94 to 100 percent of cases and the topic matched in 80 to 96 — but the source actually supported the claim only 39 to 77 percent of the time. Worse, factual grounding fell by about 42 percent as the agent ran more and more searches, while the surface metrics stayed above 92 percent. The lesson is uncomfortably compact: a working, on-topic link is not evidence.

The four-step check

The four steps get more expensive in order. Always do the first two; the third and fourth matter wherever the text leaves your hands.

1. Does it exist? For every reference, find the identifier — DOI, PubMed ID, case number, URL — and resolve it in the source's own registry. This step is near-deterministic, fast, and it removes the obvious invention. Search on the identifier rather than the title: a title can be real while the author and year are not.
2. Does it say that? Open it and find the sentence the claim rests on. This is the hard step, and the one neither tool nor model solves reliably: by a 2026 measurement, the best models caught only 37 to 61 percent of "right case, wrong pinpoint page" errors, and in two thirds of wrong acceptances they even manufactured a rationale claiming the source "expressly states" something that is not in it.
3. Is the number right? Recompute every figure and check the year and the unit. This has its own section below, because numbers fail differently from citations.
4. What do you most want to be true? Find the one claim in the text that best supports your argument, and check that one twice. This step is not about the model but about you: the figure that stays unchecked is the convenient one.

How long it takes

For an article with eight references and ten figures, the first step takes five minutes, the second ten to fifteen, the third five. So for a piece meant for publication, the full round is twenty to twenty-five minutes. That sounds like a lot until you compare it with what a correction costs.

There is a cheaper version that is enough for most internal material: step one on everything, step two on only the three claims the piece stands on. The remaining risk is not zero, but the obvious errors are gone and it takes five minutes.

Numbers are their own chapter

With citations the questions are existence and support; with numbers they are different. Four failure types occur, and each needs separate attention.

• The arithmetic error in context. One study found models do well on standalone calculation but that "performance deteriorates substantially when computations are embedded within word problems" — and the rate of logical errors rises by up to 14 percentage points as the numbers get larger. A percentage calculation embedded in a paragraph is not the same task as the same operation on its own.
• The wrong year. The figure is real, it just describes the situation three years ago. This is the most common silent error, because nothing draws attention to it.
• The wrong source. The figure is real, but it was not published by the organisation the model attributes it to. This surfaces when you resolve the identifier — and only then.
• The converted unit. Percent into percentage points, billions into millions, one currency into another. These errors arise during translation and rewriting, not in the original generation.

What web search fixes and what it does not

It would be natural to assume that letting the model search the web ends the problem. Partly true: the share of wholly invented sources does fall. But in a journalistic study of 1,600 queries, all eight search-based AI tools answered incorrectly on more than 60 percent of requests, from 37 percent error at best to 94 at worst. More importantly, the tools almost never signalled uncertainty, and the paid tiers performed worse, because they gave a definite answer in doubtful cases too.

The figure measured in medicine says the same in different words: 50 to 90 percent of responses carrying citations were not fully supported by the source named, and sometimes contradicted it. So web search largely solves the existence problem and not the support problem. For when to use which kind of search, AI search or Google sets out the criteria.

Six cases worth learning from

These are real, source-confirmed incidents from 2025 and 2026. Each failed at a different point.

• The summer reading list. In a US daily's syndicated supplement, ten of fifteen books did not exist — attributed to real, living authors. The list came from an external contributor and the newsroom did not check it.
• The quote that was never said. A technology publication retracted an article because the quotations in it were produced by a chatbot in place of a blog post that genuinely existed. The writer's employment ended within two weeks.
• The machine summary in quotation marks. A major daily published an editors' note because a reporter presented an AI summary as a direct quotation from a politician. The quotation marks are what made the error serious.
• The professor who does not exist. A Danish daily ran an article quoting an expert nobody could find, because she did not exist: the newsroom's AI system was given the task without any source material, so it produced a source to fit.
• The consultancy report. A large consulting firm's government report contained fabricated citations and a quotation attributed to a judge that was never spoken. The firm refunded part of the fee.
• The book about AI and truth. A 2026 volume shipped with more than half a dozen misattributed or fabricated quotes, including a sentence credited to a well-known journalist who never said it. The author confirmed using AI tools for research and editing.

What to take away

Never leave two things to the model: checking itself, and telling you whether the source supports the claim. The first step — resolving the identifier — is cheap and near-certain, and it removes the obvious invention. The second — whether the source actually says it — is manual work, and in 2026 it is the real risk, because the link works and the topic matches. Recompute the numbers, check the year, and verify twice the one figure you most want to be true.