The Myth That AI 'Hallucinates' the Way a Person Would

When a chatbot states a fake court case or invents a citation with total confidence, we have a word for it. We say the model is hallucinating. The term has stuck so firmly that it now appears in research papers, product documentation, and regulatory filings. It is convenient, and it is also quietly misleading.

A hallucination, in a person, is a perception without a stimulus. You see or hear something that is not there. The word carries the weight of experience: a mind, a sensory world, the sense of a thing being present. A language model has none of that. It does not perceive a false fact and report it. It assembles a plausible sequence of tokens, one after another, based on statistical patterns in its training data. Sometimes that sequence is true and sometimes it is not, and the model has no internal flag telling the two apart.

A better word already exists

Some researchers prefer confabulation, borrowed from neurology, where it describes a person filling a gap in memory with an invented detail and believing it. The key difference from hallucination is that confabulation is about producing rather than experiencing. That is much closer to what a model does. It is built to generate fluent output, and fluency does not require truth. When the training data thins out, the model keeps generating anyway, and the result is a confident answer with nothing behind it.

This is not pedantry for its own sake. The language we use shapes what we expect. Call it a hallucination and you imply the model briefly lost touch with a reality it normally tracks, as if accuracy were the default and error the glitch. The truth is closer to the reverse. The model is always doing the same thing, predicting likely text, and whether that text is correct is a separate question the model is not equipped to ask.

Why it matters

The framing feeds a larger misconception, one we have circled before: the belief that these systems possess a hidden inner life, that they know things and occasionally slip. They do not. As one widely shared 2026 essay put it, when we say a system hallucinates, we implicitly grant it a capacity for experience it has never had.

None of this means the errors are harmless or that they will disappear with a better label. Confabulated medical advice or a fabricated legal citation does real damage regardless of what you call it. The point is narrower. Understanding that a model generates rather than perceives tells you why it fails, and it sets a more honest expectation: the system is not a witness occasionally misremembering. It is a probabilistic text engine that has no idea, in any meaningful sense of the word, whether what it just told you is true.

So the next time a chatbot invents a fact with a straight face, it can help to drop the word hallucination. The machine did not see something that was not there. It simply did what it always does, and this time the pattern led somewhere false.