AI hallucinations: confident, plausible, and checkable in ten minutes.

A language model generates the next likely token given everything before it. It has no database of verified facts to consult and no internal flag that separates a remembered fact from a plausible-sounding string.

A hallucination is what happens when the most likely continuation happens to be false. The sentence that follows is fluent, correctly punctuated and confidently phrased, because confidence is a property of the prose style rather than of the underlying knowledge.

That distinction matters more than any statistic about error rates. A model does not produce hallucinations at a constant background rate; it produces them where its training data was thin, outdated or contradictory, and it produces them most readily in exactly the places where specificity is required.

Almost every hallucination in a real draft falls into one of these categories, and each needs a different check.

Ask for a citation and you will get one, because a citation-shaped string is easy to generate. Author, year, journal, volume, pages, DOI: every element has a strong pattern, so the output looks researched even when none of it exists.

There is no internal signal it can report. The model's sense of difficulty is not a sense of uncertainty about the world, so a request to flag low-confidence statements produces a plausible list rather than an accurate one.

This is why hallucination is not solvable by instruction. Prompts can reduce the frequency, and retrieval can ground the output, but neither gives you a mechanism that knows when it is wrong. Only a reader with a source does that.

Fabricated references are the most damaging hallucination because they pass the surface check. The quote reads well, the journal sounds real, the year is plausible, and the format is right, so nobody looks it up until an editor, a reviewer or a customer does.

The failure compounds in academic, legal and medical work, where a citation is not decoration but the evidence itself. A paper with three invented references does not contain three small errors; it fails as a document.

The rule is blunt: every reference, quote and attribution that will be published gets opened and confirmed by a person. No exceptions for sources that sound familiar, and none for citations generated by a tool that claims to search the web.

Two people using the same model get wildly different error rates, because error rate is a property of the task as much as the system. Ask for a summary of a document you supplied and it is low. Ask for facts about a niche, a local market or anything after the training cutoff and it rises sharply.

Temperature, retrieval, tool use and prompt specificity all shift it. So does the question's shape: requests for a specific number invite invention, while requests for a method or a checklist are mostly safe.

That is why published benchmark scores are a poor guide to your own risk. The only number that matters is how often your drafts contain a false statement, and the only way to know it is to sample and check.

High-stakes domains are not more prone to error, they are less tolerant of it. A wrong price on a landing page is a support ticket. A wrong dosage, a wrong statute, a wrong safety limit or a wrong interest rate is a liability.

Regulated claims are the sharpest edge, because a plausible invention can become a compliance breach. If a draft asserts that a product is certified, approved, clinically tested or legally required, the assertion needs a document behind it before publication.

The practical version is a tier: facts that would embarrass you, facts that would cost money, facts that would cause harm. Check everything in the third tier against a primary source, every time, without trusting the draft's confidence or a previous draft's version of the same claim.

Grounding a model in supplied documents is the single most effective mitigation available. When the answer has to come from text you provided, invented entities largely disappear and the failure mode changes to something more tractable: misreading the source.

Tractable is the right word. A model can still attribute a sentence to the wrong paragraph, invert a condition, or quietly drop a qualifier when summarising. Those errors are findable by comparing the output to the input, which is a much cheaper operation than verifying a claim against the world.

Web-connected models shift the risk again. They can retrieve real pages and still misreport them, and they cite pages that exist while misstating what those pages say. The existence of the URL proves nothing about the sentence attached to it.

Read the draft once and mark every proper noun, number, date, quotation and causal claim. That list is your verification workload, and it is usually shorter than it feels.

Check the entities first. Does the company exist, does the product exist under that name, does the person hold that role. Then the numbers, against a primary source rather than a summary. Then the quotes, verbatim, with the link kept.

Search for the sentence. A hallucinated statistic often returns nothing, while a real one usually returns a source. Absence of a result is not proof, but a result that does not match the wording you were given is a strong prompt to look closer.

Finally, delete or soften anything you cannot verify in the time available. An unsourced claim is not neutral: it reads as researched, which means it borrows credibility it has not earned.

Ask for the answer in a shape that exposes uncertainty: a table with a source column, a list where every item carries a link, or a split between what the text supports and what it does not.

Say explicitly that missing information should be marked missing rather than filled in. It reduces invention without eliminating it, and it makes the gaps visible to whoever reads the output next.

Supply the facts you already have, even if they seem obvious. Each token of supplied material is a token the model does not have to guess at, and guesses are where the damage comes from.

The most dangerous hallucination is the well-written one. A clumsy invented sentence gets caught by an editor; a smooth invented sentence gets published, quoted, and repeated by other people who assume the first writer checked.

So make checking a step in the process rather than a virtue of the person doing it. A marked-up list of claims with a source beside each one converts an unbounded worry into a finite task.

That is the whole discipline. Generation is fast and unreliable in a specific, predictable way, so verification has to be fast and systematic in the same places.