An AI hallucination is a confident, fluent answer from an AI tool that is wrong, invented, or unverifiable, stated with no signal that anything is off. It is not a glitch, a typo, or a moment where the system visibly struggles. That is what makes it dangerous: a hallucinated fact reads exactly like a correct one, in the same tone, with the same apparent certainty, and there is no built-in warning label attached to it.
The term borrows from human psychology, but the comparison only goes so far. A person who hallucinates is misperceiving something real. An AI model that "hallucinates" is not perceiving anything at all. It is predicting which words are statistically likely to come next, based on patterns in its training data, and sometimes that prediction produces a citation, a statistic, or a name that sounds exactly right and does not exist.
Why it happens, in plain terms
A language model does not have a database of facts it looks up before answering. It generates text one likely piece at a time, based on patterns it learned during training. Most of the time this produces accurate information, because accurate information is common in the patterns it learned from. But the model has no internal mechanism that distinguishes "I am confident because this is well-documented" from "I am confident because this sounds like the kind of thing that gets written in this context." Both feel identical from the inside of the process, because there is no "inside" doing any feeling or checking at all.
This is also why asking the model "are you sure?" is not a reliable fix. It can produce an equally confident-sounding correction, or an equally confident-sounding restatement of the same invented fact, because the underlying process that produced the error the first time has not changed.
The forms hallucinations actually take
Hallucinations are not always dramatic or obviously false. In practice they show up as several distinct, easy-to-miss patterns.
- Invented citations. A study, article, or report that does not exist, attributed to a real-sounding author or organization. See why AI invents sources and citations for how this specific failure mode works.
- Misattributed real sources. A real study cited correctly by name, but with a claim attached to it that the study never actually makes.
- Plausible but wrong numbers. A statistic that sounds like the kind of number that circulates in a given industry, with no real source behind it at all.
- Confidently wrong dates and versions. A claim stated as current when it describes something outdated, without any hedge acknowledging the model's training data has a cutoff.
- Filled-in specifics. A name, title, or company detail added to make an answer feel complete, when the underlying source did not actually specify it.
Why this differs from a normal writing mistake
A tired or under-informed human writer usually signals uncertainty somehow: vague language, a caveat, a note to double check something later. A hallucination carries none of that. The fluency of the sentence and the accuracy of the sentence are two completely separate things, produced by the same process, and nothing in the output distinguishes one from the other. That is the single fact this entire site's approach to AI content is built around: fluency is not evidence.
What actually reduces hallucination risk
Prompting techniques can lower the rate of hallucinations, but none of them eliminate the possibility, and none of them replace verification. Asking a model to cite sources, to say "I don't know" when uncertain, or to stick closely to a provided document all measurably help. None of them are a guarantee, because the underlying prediction process producing the text has not fundamentally changed. Treat these techniques as reducing how much verification work you have ahead of you, not as removing the need for it.
Why the term itself is a little misleading
"Hallucination" borrows a word from human perception, and some researchers argue that borrowing understates the problem rather than overstates it. A human hallucination is a misperception of something the person is genuinely trying to perceive accurately. An AI model producing a fabricated citation is not misperceiving anything, because there is nothing being perceived in the first place. It is generating the statistically likely next piece of text, with no underlying attempt at accurate perception to have gone wrong. Some writers on this topic prefer terms like "confabulation" for exactly this reason, since it describes filling a gap with a plausible-sounding invention rather than misreading something real.
The label matters less than the practical consequence: whatever you call it, the output looks the same to a reader, and the response to it is the same regardless of which term you use.
How this shows up differently depending on the task
Hallucination risk is not evenly distributed across every kind of request. A model asked to summarize a document you provided directly is anchored to real text and tends to hallucinate less, though not zero, especially on details not explicitly present in that document. A model asked an open-ended question with no supplied source material is working purely from patterns in its training data, with nothing to anchor against, and the risk rises accordingly. This is a useful practical distinction: the more a task depends on the model recalling or inventing detail from nothing, rather than working from material you gave it directly, the more verification that task deserves.
The one habit that actually matters
Every specific, checkable claim in an AI draft, whether it looks risky or not, gets treated as unverified until you confirm it yourself. This is not paranoia. It is the direct, practical response to the fact that a hallucinated sentence and an accurate sentence are, from the reader's side, indistinguishable without independent checking. The full process for doing that check, in order, is covered in how to fact-check anything an AI writes for you.
FAQ
Is an AI hallucination the same as a bug?
No. A bug is a defect a developer can typically find and fix. A hallucination is closer to an inherent property of how these models generate text: they predict likely-sounding language rather than retrieving verified facts, so some rate of invented detail is expected behavior, not a malfunction to be patched away entirely.
Do newer or larger models hallucinate less?
Newer models generally hallucinate less often on well-covered topics, but "less often" is not "never," and the difference is easy to overstate without a specific, dated benchmark behind it. Treat any claim about a model's hallucination rate as something to verify against the vendor's own current documentation before repeating it, since this changes with each release.
Can I tell a hallucination is happening just by reading the text?
Not reliably. That is the entire point: a hallucinated sentence is written in the same fluent, confident register as an accurate one. The only reliable method is checking the specific claim against an outside source, not reading the sentence more carefully.
Does this only affect chatbots?
No. Any generative AI tool that produces text, including ones built into writing assistants, search summaries, and business software, can hallucinate. The underlying prediction process is the same regardless of the interface wrapped around it.
Is a hallucination the same thing as bias in AI output?
They are related but distinct. Bias refers to systematic skew reflecting patterns in training data, such as favoring certain viewpoints or demographics. A hallucination is a specific invented or wrong detail presented as fact. A model's output can suffer from either, both, or neither in a given answer.