What Is an AI Hallucination? Meaning, Causes and Examples

In one sentence: an AI hallucination is when a language model states something false with exactly the same fluency and confidence it uses for something true. Nothing crashes, no error is raised, and the wrong answer is indistinguishable in style from a right one. The term covers invented facts, invented sources, misattributed quotes, confidently wrong numbers and — in AI agents — actions the model narrates but never performed.

This is the reference page. It covers what the word means and why it is slightly misleading, why models do it at all, why ChatGPT in particular does it, the documented cases that made the term famous, whether it is still a problem in 2026, how to tell whether an AI you rely on is doing it, and what can honestly be done about it. If you have already recognised the problem in a feature you run, the two pages that follow this one are how to cut the rate in production and how to put a number on it.

What “hallucination” means in AI — and why the word is slightly wrong

In everyday use a hallucination is perceiving something that is not there. Borrowed into AI, it describes output that is not grounded in anything: no source, no retrieved document, no fact in the training data supports it, yet the model produced it as though one did. Researchers sometimes split this into intrinsic hallucination, where the output contradicts the source it was given, and extrinsic, where the output cannot be verified against any source at all. In a product, both look the same to the customer.

The word is imperfect for two reasons worth understanding, because they shape how you fix it. First, it implies the model usually perceives the truth and occasionally sees things — but a language model never perceives truth; it produces plausible continuations, and truth is a property those continuations happen to have most of the time. Second, it implies a malfunction, when the mechanism producing the false statement is the same mechanism producing the true ones. There is no hallucination module to switch off. That is why the fixes are about grounding, constraint and measurement rather than a patch.

Why does AI hallucinate?

A large language model is trained to predict the next token — roughly, the next word-piece — given everything before it. Every sentence it produces is a continuation that is likely, given the patterns in its training data. The training objective never asks whether the continuation is true; it asks whether it is probable. Truth and probability agree often enough to make the model useful, and come apart exactly where you would expect: specific names, dates, figures, identifiers, citations, and anything that happened after the training data was collected.

Three consequences follow. The model has no internal store of facts it can consult separately from its habit of producing text, so it cannot check itself against a ledger it does not have. It has no reliable representation of its own uncertainty, so it does not sound less sure when it is guessing. And it is completing a shape as much as answering a question: asked for a reference, it produces something reference-shaped, because that is what the context calls for. OpenAI’s own 2025 analysis, Why language models hallucinate, adds a training-incentive explanation on top of the mechanical one: when evaluations score an answer as simply right or wrong, a confident guess scores better on average than an honest “I don’t know”, so models are rewarded for guessing.

Why does ChatGPT hallucinate?

For the same underlying reason as every language model, plus three things specific to a general-purpose chat assistant. It usually has no grounding: unless it searched the web or you pasted the source, it is answering from patterns in training data alone, so anything specific to your company, your documents or this week is a gap it will fill from the distribution. It has a training cutoff, and does not always know that a question falls after it. And it has been tuned to be helpful and fluent in conversation, which pushes it towards producing an answer rather than declining — the guessing incentive above, applied at the point of use.

This is also why the same question can produce a confident wrong answer one day and a correct one the next. The model has not learned anything in between; it sampled a different plausible continuation.

Examples of AI hallucinations

Three widely reported cases show the range, and each maps onto a different cause. All three were covered extensively in the press and in public legal records; the outline here is deliberately kept to what those records say.

• The fabricated legal citations (United States, 2023). A lawyer submitted a court filing containing case citations produced by ChatGPT. The cases did not exist. The model had produced citation-shaped text — plausible names, plausible reporters, plausible years — because a citation is what the request called for. The court sanctioned the lawyers involved. Cause: no grounding; the model was asked for facts it had never been given and had no way to check.
• The airline chatbot and the bereavement fare (Canada, 2024). A customer asked an airline’s website chatbot about bereavement fares and was told he could apply for the discount after travelling. The airline’s actual policy said the opposite, and was available elsewhere on the same site. A tribunal held the airline responsible for what its chatbot said and rejected the argument that the chatbot was a separate entity. Cause: retrieval that missed — the right policy existed and the system answered without it, with the added authority of being the airline’s own bot.
• The launch demo with the wrong telescope fact (2023). A promotional example for a major AI assistant stated that a particular space telescope had taken the first pictures of a planet outside our solar system. It had not; the first such images were taken years earlier by a different observatory. Cause: specificity without a source — a confidently precise claim about a fact the model had no reliable representation of, in a demo nobody fact-checked before publishing.

Eight documented cases, each traced to its cause, are collected in AI hallucination examples. The pattern across all three is the one that matters for anyone running an AI feature: nothing failed visibly, and the first person to notice was someone outside the team. The four places hallucinations enter a production feature gives the full set of causes and what to do about each.

Does AI still hallucinate in 2026?

Yes, as of September 2026, and it will for as long as the underlying architecture is what it is. What has changed is the rate and the surface. Frontier models hallucinate less often than their predecessors on benchmark tasks; systems that retrieve source material before answering do markedly better than ones that answer from memory; and structured output, tool use and verification steps have removed whole classes of invention in well-built products. What has grown is the agentic surface: an AI agent that chooses tools, calls them and narrates results can be wrong in the middle of a task in ways a single-turn chatbot never could, and that failure does not appear in the final answer.

So the honest 2026 answer to “does it still happen?” is: less, differently, and still invisibly — which is why the question that matters for a team running AI is not whether the model hallucinates but how often yours does, on your traffic. Published leaderboards measure models on benchmark tasks. They cannot tell you the rate for your feature, on your users’ inputs, against your definition of correct.

How do you know if your AI is hallucinating?

You usually cannot tell from the output alone, which is the whole problem. But there are signals worth watching for, roughly from cheapest to most reliable. The full set of checks, for readers and for teams running a feature, is in how to tell if an AI is hallucinating:

• Specificity without a source. Precise names, figures, dates or quotes that arrive with no indication of where they came from. Ask for the source; a model that has one can usually point to it.
• Citations that do not resolve. References, URLs, case names or product identifiers that look right and do not exist when checked. This is the single most reliable tell.
• Uniform confidence. The same tone for a well-known fact and an obscure one. Real knowledge is uneven; a model that never sounds less sure is not reporting its uncertainty.
• Instability on rephrasing. Ask the same thing three ways. A grounded answer survives; a hallucinated one shifts.
• In agents, narration that outruns the log. The agent says it checked, sent, updated or found something. The tool log says otherwise. This is only visible if you inspect the trace, not the reply.
• A measured rate. None of the above scales. The only reliable way to know how often a feature hallucinates is to score it against a set of real cases where you already know the right answer — which is what LLM evaluation does.

Can AI hallucinations be fixed?

Reduced substantially; eliminated, no — not with current architectures, and a vendor who promises zero is describing a demo. The interventions that actually move the rate are, in order of leverage: grounding the answer in source material and making that mandatory; fixing retrieval before touching the prompt; constraining the output shape so whole classes of invention become impossible; giving the model an explicit, licensed way to say it does not know; adding a verification pass where the stakes justify the cost; and only then changing the model. Telling the model not to hallucinate barely helps, and lowering the temperature makes a wrong answer consistent rather than correct. Each of those is a trade-off with a cost, and choosing between them requires knowing your current rate. The full breakdown, with the reasoning for the order, is here.

What this means if you run an AI feature

Everything above is true of models in general. If you have an agent, a prompt chain or a RAG pipeline serving real users, the general truth is not the useful one. The useful one is a number: on your traffic, against your definition of a correct answer, how often is it wrong — and does that number move when you change a prompt, a model or a retrieval step?

That number does not exist in any benchmark, because a benchmark measures a model and you need to measure a feature. Producing it means deciding what “correct” means for your case, decomposing that into things you can actually test, building the cases and the scoring, and then running it as a gate every time something changes. Hallucination is one failure class inside that — usually the one that prompted the question — but the same measurement catches the wrong tool call, the discarded intermediate result and the answer that was right for a reason that will not hold next week. That is what we mean by LLM evaluation, and it is the point at which “the AI sometimes makes things up” stops being a complaint and becomes a metric with a trend line.

Frequently asked questions

What is an AI hallucination in one sentence?

An AI hallucination is a language model stating something false with the same fluency and confidence it uses for something true, because it was never given, or never retrieved, a source for the claim.

What is the difference between an AI hallucination and a bug?

A bug is code doing something other than what it was written to do, and it usually fails visibly: an error, a crash, an obviously malformed result. A hallucination is the model doing exactly what it was built to do, producing a plausible continuation, in a case where plausible and true came apart. Nothing fails, nothing is logged, and the output is indistinguishable in form from a correct one. That is why hallucinations are managed by measurement and grounding rather than fixed by a patch.

Why does ChatGPT hallucinate?

For the reason every language model does, namely that it predicts likely text rather than checking true text, plus three things specific to a general chat assistant: it usually answers from training data alone with no grounding in a source, it has a knowledge cutoff it does not always know it has passed, and it has been tuned to be helpful and fluent, which pushes it towards producing an answer rather than declining. The same question can therefore be answered wrongly one day and correctly the next; the model sampled a different plausible continuation rather than learning anything in between.

Does AI still hallucinate in 2026?

Yes. As of September 2026 frontier models hallucinate less often than earlier generations on benchmark tasks, and systems that retrieve source material before answering do markedly better than ones answering from memory, but the mechanism is architectural and the rate is never zero. The surface has also shifted: AI agents that choose and call tools can be wrong in the middle of a task in ways a single-turn chatbot could not, and those failures do not appear in the final answer. The practical question is no longer whether models hallucinate but how often a specific feature does, on its own traffic.

How do you know if your AI is hallucinating?

Usually not from the output alone. The signals to watch are precise claims that arrive with no source, citations or identifiers that do not resolve when checked, uniform confidence across easy and obscure questions, answers that change when the question is rephrased, and, in agents, narration that says something was checked or done when the tool log says otherwise. None of these scale. The reliable way is to score the feature against a set of real cases where the correct answer is already known, which gives a measured rate rather than an impression.

Can AI hallucinations be fixed?

They can be reduced substantially but not eliminated with current architectures. The interventions that move the rate, in order of leverage, are grounding the answer in source material, fixing retrieval before touching the prompt, constraining the output shape, giving the model an explicit way to decline, adding a verification pass on high-stakes paths, and only then changing the model. Instructing the model not to hallucinate barely helps, and lowering the temperature makes a wrong answer consistent rather than correct. Because each intervention has a cost, choosing between them requires knowing the current rate on your own traffic.

Is "hallucination" the right word for it?

It is the established term and it is imperfect. It implies the model normally perceives the truth and occasionally sees things, when in fact a language model never perceives truth at all; it produces plausible text, and truth is a property that text has most of the time. It also implies a malfunction, when the mechanism producing a false statement is the same one producing true ones. Alternatives such as confabulation or fabrication are sometimes used. Whatever the word, the operational definition is the useful one: a confident output with no source behind it.