Glossary
Hallucination
A hallucination is an AI system stating something false as though it were established fact, such as an invented source or a wrong figure, produced in the same fluent tone as its correct answers.
In plain terms
It is the confident wrong answer. Not a typo and not a refusal, but a smooth, well-formed statement that happens not to be true: a court case that was never heard, a feature a product does not have, a date that is a year out. The unsettling part is that it reads exactly like the parts that are right, because nothing in the writing marks the difference.
Why it matters
This single behaviour decides how these tools can safely be used at work. It is why an assistant is excellent at drafting and unreliable at recalling. It is why the useful question is what you can check rather than what it can produce, and why the tasks that pay off are the ones where you already know enough to spot a wrong answer. Teams that understand it end up with a sensible division of labour, letting the software do volume and structure while people keep the facts. Teams that do not either trust it too far and get caught, or distrust it entirely and get nothing.
How it works
The mechanism is the same one that produces correct answers. The software generates a plausible continuation of the text it has been given, and plausibility is judged from patterns in what it read, not against any record of what is true. A well-formed invention and a well-remembered fact look identical from the inside, because the process that produced them is identical.
The failure clusters where the training material was thin, contradictory or absent. Ask about a widely documented subject and the plausible continuation is usually also the accurate one. Ask about a niche regulation, a small company, a recent event or a specific person, and there is no dense pattern to draw on, so the most plausible-sounding text gets produced anyway. Specific, checkable details are the highest-risk category: names, numbers, dates, citations and quotations.
Refusing is a learned behaviour rather than a natural one. Models are trained after the fact to say they do not know, but that training competes with the pull towards producing a helpful answer. Push hard for a specific answer and you increase the chance of getting an invented one, because the request itself has raised the cost of admitting ignorance.
Supplying the material changes the odds substantially. When the answer is in text the model can actually see, the plausible continuation and the accurate one converge, which is why retrieval and document upload reduce the problem so effectively. They do not remove it. A system reading your documents can still misread them, over-generalise from them, or fill a gap between two passages with something that sounds like it belongs.
There is no internal confidence signal you can read from the writing. The tone is a property of the training rather than a measure of certainty, so hedged phrasing does not reliably mark the shaky parts and firm phrasing does not mark the sound ones. Asking a model how sure it is produces another plausible continuation rather than an audit.
The most reliable defence is choosing tasks where verification is cheap. Where you supply the material and ask for it to be restructured, summarised or rewritten, the facts came from you and the risk is largely confined to misreading. Where you ask for facts you do not have, every specific detail is something you must check, and the effort of checking can exceed the effort of writing. The distinction is worth applying deliberately when deciding what to hand over.
Watching for the shape of the failure helps more than watching for a feeling. Invented material tends to be suspiciously well-formed: a citation with a plausible journal, volume and year, a statistic quoted to a decimal place, a quotation that reads better than the source usually does. Precision without a route back to the source is the signal worth reacting to, and it is one you can check in seconds.
Same question, with and without the material
Seen in the wild
Ask a general assistant for sources on a niche subject and check that each one exists. Invented references are the classic case, because the shape of a citation is easy to imitate.
ChatGPTAsk a research assistant the same question and follow its numbered links back to the pages they came from, which turns a claim into something you can check in one click.
PerplexityLoad your own documents into a notebook tool and ask a question the material does not answer, then watch whether it says so or reaches beyond what you loaded.
NotebookLM
Common misconceptions
People assume
Hallucination is a bug that will be fixed in the next version.
In fact
It is a consequence of how the software works rather than a defect bolted on to it. Newer models hallucinate less on well-covered subjects, and grounding techniques cut it further, but nothing on offer removes it. Plan for a system that is sometimes confidently wrong rather than waiting for one that never is.
People assume
It only happens on obscure questions.
In fact
Obscure questions are the highest-risk category, but the failure appears anywhere the training material was thin or inconsistent, including recent events, small organisations, specific figures and anything involving a precise date. Familiar subjects are safer, not safe.
People assume
You can tell from the wording when it is unsure.
In fact
You cannot. Tone is a product of training rather than a measure of confidence, so inventions arrive in the same fluent register as facts. Asking how certain it is produces another plausible answer rather than a genuine assessment.
Telling them apart
Hallucination vs Citation
Hallucination
A claim with nothing behind it. Reads the same as a sound one, and the only way to tell is to go looking for the source yourself.
A claim with a route back to where it came from. Turns verification from a research task into a click, which is why systems that cite are easier to trust than systems that merely sound confident.
A citation is not proof the claim is right, only that a source exists and can be checked. An invented citation is the classic case, so the check is following the link rather than seeing that one is offered.
Hallucination vs Grounding
Hallucination
The failure: stating something that is not so, in the same tone as everything else, because the material to answer from was not there.
The main countermeasure: supplying real source material with the question so the answer is drawn from it rather than from recall, usually with citations back to the passage used.
They are two ends of the same problem. If you find yourself asking whether to trust an answer, the fix is almost always to give the system the documents rather than to ask more carefully.
Questions
- Why is it called hallucination?
- By analogy: the system produces something that is not there, with the same conviction as things that are. The word is disputed, because it suggests a perceptual error rather than the actual mechanism, and some researchers prefer confabulation or fabrication. The term is established in general use, so it is the one you will meet.
- Which tasks are most affected?
- Anything hinging on a specific verifiable detail: citations, statistics, dates, legal or regulatory specifics, quotations, and facts about small organisations or recent events. Tasks where you supply the material and the model transforms it, such as summarising, rewriting or reformatting, are considerably safer because the facts come from you.
- Does giving it my documents stop it?
- It reduces it substantially and does not stop it. With relevant text in front of it, the plausible answer and the accurate one usually coincide. The residual risk is misreading a passage, over-generalising from one, or bridging a gap between two with something invented, so citations that let you check the source still matter.
- How do I check an answer efficiently?
- Check the checkable parts rather than re-reading the whole thing. Verify names, figures, dates and any cited source, because those are where invention concentrates and where a single wrong item does the most damage. Structure, reasoning and prose quality rarely need the same scrutiny, and treating them all alike is what makes verification feel expensive.
- Do newer models still do this?
- Yes, though less on well-covered subjects. Each generation reduces the rate rather than eliminating the behaviour, because it follows from how the software produces text at all. Treat improvement as a reason to widen what you use these tools for, never as a reason to drop verification on details that matter.
- Should this stop us using AI at work?
- No, but it should shape what you use it for. The good applications are ones where you can verify the output or where being occasionally wrong is cheap: drafting, summarising material you supply, restructuring, brainstorming. The poor ones are those where an unnoticed wrong fact reaches a customer, a regulator or a contract.
Key takeaways
- A hallucination is a confident false statement produced by the same process that produces correct ones.
- Risk concentrates on specific checkable details: names, figures, dates, citations and quotations.
- Tone carries no information about certainty, and asking how sure it is does not help.
- Supplying the source material substantially reduces it without removing it.
- Use these tools where you can verify the output, and keep the facts on the human side.
Tools that use this
- Perplexity
Answers carry numbered links back to sources, so claims can be checked directly.
- NotebookLM
Answers stay within documents you supply, which is the grounding case in its strictest form.
- ChatGPT
The general assistant where unsourced recall is easiest to test against reality.
Last checked July 2026