How to Catch an AI Hallucination Before It Reaches a Client

A law firm submitted 42 fabricated citations to a judge. A national newspaper published a quote nobody had said. A police officer is under criminal investigation over AI-generated evidence. Officials have been suspended and a scientific paper retracted.

In none of these cases did anyone act carelessly by their own account. They reviewed the work. The output looked right — and looking right is precisely the thing these systems are most reliably good at, independent of whether the content is true.

That's what makes hallucination different from ordinary error. A tired human produces work that shows signs of being rushed. A language model produces its errors in the same confident register as its correct answers, with the same formatting and the same air of having checked.

The one distinction that predicts most failures

Did the model have the source, or is it recalling?

Given a document, a model is reading. It's genuinely good at that: extracting, comparing, summarising, cross-referencing. Errors here are usually omissions — a lost qualifier, a dropped caveat, a flattened confidence level.

Without the document, it's reconstructing from training data, and the quality of that reconstruction is unknowable from the output. This is where invented citations come from. Not sloppiness — a plausible reconstruction of what a paper on that topic, by that author, in that year, would have been called.

Structure your work so the model is almost always reading. When it can't be, treat what comes back as a list of leads rather than a set of facts.

The four things that hallucinate most

Citations and sources. Anything with an author, a title, and a year attached. The failure is confident and specific, which is what gets it past review.

Numbers. Especially aggregates, percentages, and anything requiring a choice of denominator. A total that's slightly off is invisible in a clean table.

Quotes. Composites of things a person plausibly said, in their register. Almost impossible to spot on internal evidence alone.

Dates and currency. A model is frequently vague about its own knowledge cutoff. Anything stated as current deserves an "as of when?" — and scepticism about the answer.

A verification pass that takes four minutes

Not "read it carefully" — a checklist, because the whole problem is that careful reading doesn't catch this.

1. Open every citation. Not scan — open. And check that it says what the summary claims, which is a separate and more common failure than the source not existing.
2. Trace one number per section back to its origin. One, not all. It catches most of what matters.
3. Search any direct quote verbatim. Ten seconds each.
4. Ask "as of when?" for anything presented as the current state of something.
5. Name the mechanism behind any causal claim. If you can't state why A would cause B, you have a pattern, not a finding.

Four minutes, and it's shorter than the correction email.

What doesn't work

Asking the model to check itself. It will confirm its own output with the same confidence it produced it. Self-verification is the appearance of a check.

Using one AI to fact-check another. Better than nothing, and still not a check — models share training data and failure modes, and both can be confidently wrong in the same direction. Useful for finding disagreement, which flags something worth your attention. Not useful as a verdict.

Reading for plausibility. That's the exact filter hallucinations are optimised to pass.

Tomasz — the trusted author

Tomasz is a policy researcher at a think tank. He checked the sources in an AI-drafted background section and found them fine, except one he skimmed because the author was someone he'd read before.

The paper didn't exist. Real author, plausible topic, credible year, exactly the title such a paper would have. It reached internal review before a colleague tried to download it.

His rule since: every citation gets opened, without exception, especially the ones from authors he trusts. Familiarity was the thing that let it through.

Rina — the number that survived three reviews

Rina is a compliance manager at a financial services firm. A regulatory summary went to 40 client firms with a figure that was accurate for the prior reporting period and had since changed.

Three people had reviewed it. All three read for plausibility, and the number was entirely plausible — it had been correct four months earlier.

She now requires one figure per section to be traced to source before anything goes out. It adds a few minutes. It has caught two similar cases since, both of them things that were true once.

The one thing

Hallucinations don't announce themselves, which means no amount of careful reading catches them. Only checking against something outside the document does.

Open the citation. Trace the number. Search the quote. Ask when. It's four minutes, and it's the difference between a routine deliverable and the story your profession tells about you.