An AI hallucination is confident output from an AI tool that is not grounded in fact or in the information you gave it — the model states something as true that is invented, wrong, or unsupported, and it does so in the same fluent tone it uses when it is right.
The honest answer on frequency is that there is no single reliable rate: it varies enormously by model, task, and how the tool is used, and the reference material for this page contains no frequency or percentage data at all, so any specific number you see quoted should be treated with suspicion unless the source explains exactly what was measured and how.
The mechanism behind hallucinations is easier to understand once you stop thinking of the model as a database. A large language model is trained to predict the next likely chunk of text given everything before it. That training produces something that behaves like a very well-read pattern matcher.
When you ask a question, it does not look up a stored answer and hand it back. It generates the most plausible-sounding continuation. Most of the time, plausibility lines up with truth because the training data contained a lot of true statements.
But plausibility and truth are not the same thing, and the gap between them is where hallucinations live. This is why a model can invent a citation with a real author and a fake page number, or describe a feature in a product that has never existed — both are plausible continuations, neither is grounded.
It is also why the same model that nails a summary of a well-documented topic can fall apart on a niche question where the training signal was thin.
A concrete example makes this clearer. Suppose you ask a tool to summarise a 12-page contract you paste in. You ask: "What is the notice period for termination?"
The contract says 60 days in clause 9. If the model answers "60 days, per clause 9," that is grounded — it pulled from your input. If it answers "30 days, per clause 4," it has hallucinated: clause 4 might exist, 30 days might be a common default, and the whole answer reads perfectly.
The tell is not the tone. The tell is that the number does not appear in the document you supplied. This is the single most useful habit you can build: for any factual claim, ask "where did this come from — my input, or the model's memory of the world?" If the answer is "memory," treat it as a claim to verify, not a fact.
So how often does it happen? Honestly, the reference material gives no rate, and the real-world answer depends on factors that no single statistic captures. Short factual questions about widely covered topics tend to be more reliable.
Long chains of reasoning, niche or recent topics, precise numbers, and requests to cite sources are where hallucinations concentrate. A model asked to write a poem about the ocean is effectively never "wrong" in a checkable way — there is no ground truth to violate. A model asked for the population of a small town in 1998 has a specific, checkable claim to get right, and much more room to fail.
That is why the useful question is not "what percentage of AI answers are wrong" but "for this task, on this topic, how much can I verify?"
Two practical rules follow. First, separate tasks into low-stakes and high-stakes. Drafting a brainstorm list is low-stakes; a legal, medical, or financial claim that will affect a decision is high-stakes, and every specific number, name, date, or citation in it should be checked against a primary source.
Second, give the model the source material and ask it to answer only from that material. Grounding the answer in a document you control removes most of the room for invention, because the model is now summarising rather than recalling. According to our AI tool database, we maintain snapshots of 360 AI tools with pricing and capability details recorded at verification time, and the most recent verification date is 2026-09-18 — a reminder that even tool facts go stale and should be checked rather than assumed.
If you want to go deeper on why models invent things and how to spot it, the guide on why AI sometimes makes things up is a good next read.