Part of our guide to How to Check an AI Answer
"Hallucination" is a borrowed word and a slightly misleading one. Nothing is malfunctioning. The system is doing precisely what it does, on a question where that is not enough.
A model produces text that fits. When it holds the fact, the fitting text is true. When it does not, an invented but plausible answer fits equally well — and nothing in the process marks the difference. It is not lying; there is no internal fact being concealed.
Why it happens
The model was trained to continue text convincingly. It was not built as a database with a lookup that can fail loudly and return nothing.
So when a question requires a detail it does not hold — a specific figure, an exact citation, a date — there is no "I do not have this" state to fall into. The most probable continuation is a confident, well-formed answer, and it produces one.
This is why the errors cluster on specifics. General explanations are supported by enormous amounts of consistent training text. A single obscure number is not.
What makes it worse
Obscurity. The less something appeared in training data, the more the model is filling gaps.
Requests for precision. Asking for exact figures or citations invites invention.
Pressure to answer. A question framed as though a definite answer must exist tends to produce one.
Long chains of reasoning. An early invented step gets treated as established and built upon.
Stop asking the model to recall and start asking it to work with. Paste the document, the page, the data. Grounded in supplied material, the failure rate drops sharply — because the task changes from remembering to reading.
Where it shows up most
Not evenly. The failure clusters in predictable places, which is what makes it manageable rather than paralysing.
| Most likely to be wrong | Why |
|---|---|
| Citations, references, DOIs | Their surface form is highly patterned and easy to generate convincingly |
| Precise figures, dates, statistics | Plausible-looking numbers are cheap to produce; correct ones require retrieval |
| Quotations attributed to people | Style is imitable, provenance is not |
| Obscure or specialist facts | Sparse training coverage, but no signal that coverage was sparse |
| Recent events | Beyond the training cutoff, and the model may not flag that |
| Legal and regulatory specifics | Jurisdiction-dependent, and change often |
| Less likely to be wrong | Why |
|---|---|
| Widely documented general concepts | Represented consistently across many sources |
| Structure, format, rephrasing | No external fact is being asserted |
| Reasoning over material you supplied | The facts came from you |
| Code that you then run | It either works or it does not — the check is immediate |
The pattern underneath: error rises as the answer depends on recalling a specific fact, and falls as it depends on manipulating something you provided.
What confidence does not tell you
The most useful thing to internalise is that fluency and accuracy are unrelated in the output.
A model does not have a reliable internal signal of "I do not know this" that it surfaces in its wording. A fabricated citation arrives in exactly the same measured, well-formed prose as a correct one. There is no tell — no hedging, no hesitation, no drop in quality.
This is genuinely different from talking to a person, where uncertainty usually leaks into how something is said. Human intuition about confidence is well tuned and it does not transfer, which is precisely why people who are careful with sources still get caught.
Asking "are you sure?" does not fix it either. The follow-up is generated the same way as the original, and a model will sometimes revise a correct answer under pressure just as readily as an incorrect one.
What it means for how you use it
Treat model output as a draft to verify, not a source to cite. It is genuinely useful for explanation, structure, summarising material you provide and generating options.
It is not a reference work, and the fluency of the writing is not evidence about the accuracy of the content. Those two things feel connected and are not.
Related reading
This is general information — see our disclaimer.
Frequently asked questions
Is hallucination a bug?
Not in the usual sense. The system generates plausible continuations of text; when it lacks the fact, a plausible invention fits the pattern as well as the truth would.
Can it be fixed completely?
It can be reduced substantially — by grounding answers in supplied documents or search results — but a model generating from memory alone will always be able to produce confident errors.
Does a bigger model hallucinate less?
Generally somewhat less on common facts, but not zero, and larger models can be more persuasive when wrong. Scale is not a cure.
How do I reduce it in practice?
Give the model the source material rather than asking it to recall, ask for uncertainty to be stated, and verify anything specific.
Sources
Found an error? Email us and we will fix it and note the change at the bottom of this article. Hello@daily-atlas.com
