What is an AI hallucination?

A large language model is trained to predict the next piece of text. It learns which words tend to follow which, not a database of verified facts. When you ask about something it learned well, the likely answer is usually the right one. When the question touches a gap in its training data, a rare detail or something that happened after its training ended, it still produces the most plausible-sounding text, because that is what it was built to do.

Think of a student who never leaves an exam answer blank: when unsure, they write something that looks like a real answer. A model has no built-in sense of 'I am not sure' unless it has been trained or prompted to express doubt, and it writes errors in the same fluent, confident tone as facts.

The US National Institute of Standards and Technology (NIST) lists this risk as 'confabulation' in its generative AI risk profile, a term that avoids implying the model perceives anything. Researchers also separate output that contradicts the source the model was given from output that adds claims the source cannot support (Ji et al., 2022). In everyday use, the most frequent forms are:

A hallucination is not a lie: the model has no intention to deceive and no awareness that it is wrong. It is also different from outdated knowledge (information that was correct but has since changed) and from bias (a systematic slant learned from training data), although all three can appear in the same answer.

Nor is it a rare glitch that one update will remove. It follows from how these models generate text, so every AI tool can do it, including those that search the web or read your files.

You cannot switch hallucinations off, but you can make them rarer and easier to catch. When you compare AI tools, look at whether answers show their sources and whether each claim can be traced back to them; that says more about reliability than a general accuracy claim. Practical habits: