AI Hallucinations and How to Catch Them

A hallucination is a confident, fluent statement that is not true, delivered in the same tone as everything else the model says. It is the most expensive failure mode of language models, because it stays invisible until somebody acts on it. Most hallucinations can be caught with cheap habits, provided you know where to look.

Why a model invents things

The model predicts likely text, and a citation that looks like every other citation is likely text. It has no internal ledger sorting what it absorbed from what it assembled. When it needs a source for a claim, it produces a string that fits the pattern: author, year, journal, volume, pages, identifier. That string can be complete fiction while looking exactly like a real record.

Training data also has gaps and stale edges. A model asked about a niche topic may merge two real entities, attach one person's achievement to another, or describe a field as it stood three years ago. Confidence is unrelated to accuracy, so asking whether it is sure changes very little.

The shapes hallucinations take

Citations that do not exist, or that exist with the wrong authors. Legal cases with plausible names and no record in any reporter. Quotations attributed to real people who never said them. Statistics quoted to a precise decimal that no study contains. Biographies that fuse two people with the same name. Software functions that were never written. Policies from an organisation that has no such document. Sums of money and dates shifted by a plausible margin.

The common thread is plausibility rather than randomness. Invented material is not strange, it is familiar in every respect except correspondence to reality, which is precisely why it survives a quick read. A fabricated case name reads like a case name because it was assembled from the statistical shape of thousands of real ones.

Warning signs in the text

Watch for a source that cannot be found when you search its exact title. Watch for hedging that vanishes, where a model says usually and sometimes in one paragraph and states a hard number in the next. Watch for answers that reshape themselves to match your question, which points to agreement rather than knowledge. Watch for perfect specificity about a small, obscure topic, because that is where pattern matching has the least real material to work from.

Be suspicious of any quote you cannot locate, any number you cannot trace, and any claim that would be surprising if true. Surprising true claims leave a trail of records and coverage. Invented ones leave nothing.

How to verify quickly

Paste the exact title of any citation into a search box. An invented paper usually returns nothing but the sentence you pasted, sometimes with the model's own wording as the only match. For academic sources, resolve the identifier directly, since a real one points to a real record with the stated authors.

Ask the model for the specific sentence it is quoting and the page number. If it cannot produce one, then rewrites the quotation when pressed, treat the whole passage as fiction. A useful prompt is: For each claim above, mark it verified, uncertain or unknown, and name the basis you are relying on. That does not make the model honest, but it often exposes the parts it is least sure of.

Where hallucination does the most damage

Hallucination is cheapest where you were going to check anyway, such as brainstorming, and most dangerous where trust is automatic. Medical dosing, drug interactions, legal deadlines, tax rules, safety procedures, immigration requirements, and anything a client will act on without asking further questions.

In all of those, a model is a way to find the right questions, not a source of the answer. Go to the primary document, the regulator, or a qualified professional, and treat the model's output as a map of where to look rather than the finding itself. If you catch one invented detail in an output, assume there are others you have not checked, because errors cluster rather than arrive alone.

A checking routine that takes two minutes

• Name the claim - Write down in one sentence what the output is asking you to believe.
• Find the primary source - Look for the original document rather than another summary of it.
• Match the numbers - Check each figure against the original, including units and dates.
• Test the quote - Search the exact phrase. A quote nobody can find is not evidence.
• Ask for uncertainty - Have the model label each claim and name the basis it is relying on.
• Keep the receipt - Save the page or link you verified, so you can point to it later.