When an AI is "hallucinating," it is producing fluent, confident output that is factually wrong or unsupported by the information it was given.
The word sounds dramatic, but the behavior is mundane: the model generates text that reads perfectly while being false — an invented citation, a fake statistic, a function that does not exist, a summary of a document that never said that. The key detail is the confidence. A hallucination does not look like a guess. It looks exactly like the model's correct answers, which is why it fools people who are skimming.
Why does this happen? A large language model is a next-token prediction system. It was trained to produce the most plausible continuation of a piece of text, not to verify whether that continuation is true.
Those two goals usually overlap — plausible text is often accurate text — but they are not the same goal, and the gap between them is where hallucinations live. When the model lacks a fact, it does not have a reliable internal flag that says "I don't know." It has a strong tendency to keep generating, because fluent continuation is what it was optimized to do.
This is also why hallucinations are worse in obscure territory: ask about a famous event and the training signal is dense; ask about a niche regulation or a small library's API and the model is filling gaps with plausible-sounding structure. It is worth separating hallucination from two neighbors.
Ordinary error is when the model gets something wrong that it could have gotten right — a miscalculation, a misread instruction. Outdated information is when the model is correct about an old state of the world but not the current one. Hallucination is specifically fabrication: content with no grounding in either the training data or the prompt.
That distinction matters because the fixes are different. For outdated information, you need fresher sources. For hallucination, you need verification.
Here is a concrete example you can reproduce. Ask a chatbot for a peer-reviewed study on a very specific claim — say, the effect of a particular study technique on a particular age group — and ask for the author, journal, and year. You will often get a tidy answer: a real-sounding author, a real journal, a plausible year, and a title that reads like the literature.
Then search for that exact title. It frequently does not exist. The journal is real, the topic is real, the citation is invented.
The same thing happens in code: ask for a helper function in a library and the model may confidently produce `library.parseConfigFile()` with arguments that look right, when the actual function is named something else or lives in a different module. The tell is not the tone — the tone is always confident.
The tell is that the specific, checkable detail fails when you check it. A useful habit is to treat every specific number, name, date, quote, and API signature as a claim to verify, not as a fact delivered.
Now the honest limits. Hallucination cannot be fully eliminated in current systems, because the underlying mechanism — generating plausible text — is the same mechanism that makes the tool useful. Mitigations reduce the rate and the damage but do not remove the risk.
Retrieval, where the model is given documents to work from, grounds answers in supplied text and cuts down on invented facts, but the model can still misread or overstate what those documents say. Asking the model to cite sources helps you audit its claims, but a citation can itself be fabricated, so you still have to open it.
Lowering the temperature makes output less random, which can reduce some creative drift, but it does not install a truth check. The practical rule is proportional: for low-stakes drafting, brainstorming, or rewriting your own text, unverified output is usually fine. For anything that goes into a contract, a medical decision, a legal filing, a financial report, or production code, treat the model as a first-draft generator and verify every factual claim and every identifier against an external source.
According to our AI tool database, which maintains pricing and capability snapshots for 360 AI tools as of its most recent verification, capability descriptions are snapshots rather than guarantees — the same caution applies to model outputs. If you want to see how this plays out across different tools and workflows, our guide on why AI sometimes makes things up and how to tell when it's wrong walks through the practical checks.
A final insight that goes beyond the obvious: the most dangerous hallucinations are not the wild ones. A model claiming a cure for a disease is easy to dismiss. The dangerous ones are small, boring, and embedded in otherwise correct work — a slightly wrong version number, a real author attached to the wrong paper, a statistic that is off by a decimal place. Those survive review because everything around them is right. So build the verification step into the task itself, not into your mood. If a number or a name matters, it gets checked. Every time.