An AI hallucination is an answer produced by an AI system — a chatbot, a writing assistant, an image generator — that is stated fluently and confidently but is wrong: a fact that is not true, a book or study that does not exist, a date, quote or statistic invented on the spot. The output sounds exactly like a correct answer. That is the whole problem. A hallucination is not an answer that looks broken; it is an answer that looks finished.

The US National Institute of Standards and Technology (NIST) uses a blunter word for the same thing in its Generative AI Profile (NIST AI 600-1): confabulation, defined as “confidently stated but erroneous or false content” that can mislead or deceive the reader. IBM’s explainer describes AI hallucinations the same way — outputs that are plausible on the surface but factually wrong, irrelevant or entirely fabricated, including fake facts, invented studies and nonexistent links presented as real.

One distinction matters before anything else: a hallucinating model is not lying. Lying requires knowing the truth and choosing to say something else. A language model does not know anything in that sense. It produces the most likely continuation of the text it was given, and sometimes the most likely continuation is false. The model is not deceiving you on purpose — it is doing the only thing it was built to do, in a situation where that thing fails.

What counts as an AI hallucination?

Not every AI error is a hallucination. A spelling mistake, a clumsy sentence or an answer that is simply out of date is an ordinary error. A hallucination has a specific shape: the content is invented or unsupported, and delivered with the same confidence as a verified fact. The recurring forms are these:

  • Fabricated facts and citations. The model invents a statistic, a historical event, a person’s biography — or a complete academic reference with a plausible title, real-sounding authors and a journal that never published it. Fake citations are the signature hallucination because they are so easy to check and so rarely checked.
  • Conflation. Details from two real things are merged into one wrong thing: two people with similar names become one person, two products become one product, two court cases become a single case with the facts of both.
  • Contradicting the source. You give the model a document and ask for a summary, and the summary states the opposite of what the document says, or adds claims the document never made. Researchers divide this into intrinsic hallucination (the output contradicts the source) and extrinsic hallucination (the output adds material the source cannot support). You will also see the same split labelled factuality — the output clashes with the real world — versus faithfulness — the output clashes with the source it was given. Different labels, same two failures.
  • Confident nonsense on unknowable questions. Ask about an obscure person’s birthday, an unpublished figure, or something that happened after the model’s training data ends, and instead of saying it does not know, the model produces a specific, wrong answer.
  • Image hallucinations. Image models add fingers, merge objects, garble text on signs, or render physically impossible scenes — the visual version of the same behaviour.

Why do AI models hallucinate?

Three causes stack on top of each other, and you need all three to understand why the problem survives every new model release.

1. The model predicts text; it does not look facts up

A large language model is trained to predict the next word (technically, the next token) given the words before it, across an enormous body of text. That training makes it superb at producing text that reads as true, because true statements and plausible-sounding statements have the same shape. Nothing in the basic mechanism checks a claim against the world. When the model reaches a gap — a fact it never saw, or saw too rarely to retain — it does not stop. It fills the gap with whatever continuation is statistically most likely, which is how you get a perfectly formatted citation to a paper that was never written.

2. The training data has gaps, errors and a cut-off date

Models learn from text collected at a particular time. That text contains mistakes, myths repeated as fact, satire, fiction and outdated information, and it ends at a cut-off point. Questions about rare topics, very recent events, local details or a specific organisation’s internal facts sit exactly where the data is thinnest — and thin data is where hallucination rates climb.

3. Training and testing reward guessing over honesty

This is the cause a September 2025 paper from OpenAI researchers put at the centre of the problem. In “Why Language Models Hallucinate”, the authors argue that models hallucinate in part because the way they are trained and scored treats a guess as better than an admission of ignorance. On most benchmarks, answering “I don’t know” scores zero, while a guess scores zero only when it is wrong and full marks when it is right. A model tuned to maximise score learns to behave like a student who would rather bluff than leave a blank — and bluffing, stated smoothly, is precisely what a hallucination is. The paper’s proposed fix is not a bigger model but different scoring: penalise confident wrong answers more heavily than uncertainty, so honesty stops being a losing strategy.

Layered together, the three causes explain the pattern you actually see: hallucinations are rare on famous, heavily documented topics (the prediction is usually right because the data is thick) and common on obscure, recent, local or oddly specific ones.

Why AI hallucinations are a real problem

In casual use, a hallucination costs you a wrong trivia answer. In serious use, the same behaviour has produced documented harm:

  • Law. Lawyers in multiple countries have filed court documents containing case citations invented by chatbots — cases that do not exist, with quotes attributed to judges who never wrote them. Courts have sanctioned the lawyers involved. A maintained database of such incidents now runs to well over a thousand legal cases worldwide involving AI-hallucinated content.
  • Health and money. A confident, wrong answer about a medication, a symptom or a tax rule is more dangerous than an obviously bad one, because nothing in the presentation warns the reader to check. NIST lists confabulation among the defining risks of generative AI for exactly this reason, and adds a human factor on top: automation bias, our tendency to trust machine output because it is machine output.
  • Work product at scale. A hallucinated figure pasted into a report, a contract summary that reverses a clause, a product description inventing a feature — each error is small, and each is expensive to find later because it arrives dressed as finished work.

Two cases show how ordinary the failure looks from the outside. In Mata v. Avianca in 2023, lawyers in New York filed a court brief built on cases a chatbot had invented wholesale — names, citations and quotes, none of them real — and the judge sanctioned the lawyers and their firm. The brief read like any other brief. That is why it got filed. A year later in Canada, Air Canada’s own website chatbot told a passenger he could claim a bereavement discount after his trip. No such rule existed. A civil tribunal ruled against the airline and held it responsible for what its chatbot had said — the company could not shrug the answer off as the machine’s mistake rather than its own.

The OWASP GenAI Security Project treats hallucination as a security concern for the same reason: a model that invents APIs, package names or policy details can push false material into systems and supply chains, not just into essays.

How to spot a hallucination

You cannot detect hallucinations by tone — the false answer and the true one have identical tone. You detect them by checking, and some checks are far cheaper than others:

  1. Verify every citation. If an answer names a paper, book, case or article, search for it. If you cannot find it anywhere except in the AI’s answer, treat it as invented until proven otherwise. This single check catches the most damaging class of hallucination.
  2. Check names, dates and numbers individually. Hallucinations cluster in the specifics. A broadly correct explanation with one wrong date or one invented statistic is the typical shape.
  3. Ask for the source, then open it. A model asked to ground its answer may produce a real source — or another hallucinated one. The link only counts if it resolves to a page that actually says what was claimed.
  4. Cross-check with a second, independent route. Search the question yourself, or check a primary source. Two AI chatbots agreeing is weaker evidence than it looks, because models are trained on overlapping data and share each other’s characteristic errors.
  5. Be suspicious in proportion to obscurity. The more specific, local, recent or rarely documented the fact, the higher the chance the model is filling a gap rather than reporting. Famous facts are usually safe; the long tail is where you verify.

How to reduce hallucinations

Nothing eliminates them — IBM’s assessment, and the position across the research literature, is that hallucinations can be reduced but not fully removed, because they grow out of the prediction mechanism itself. What works is pressure from both sides.

If you use AI tools:

  • Give the model the source. Paste or attach the document and ask for an answer based only on it. Answers grounded in provided text hallucinate far less than answers from memory — the same reason it matters which plan limits apply to files you upload to ChatGPT when you work with long documents.
  • Ask questions the model can refuse. “If this is not in the document, say so” gives the model an exit that plain questions do not.
  • Narrow the question. Specific, bounded questions (“what does section 3 say about refunds?”) fail less often than open ones (“tell me about this company’s policies”).
  • Verify the load-bearing claims. You do not need to check every sentence. Check the ones a decision depends on.

If you build or deploy AI systems:

  • Retrieval-augmented generation (RAG) and grounding — connecting the model to a searched, verifiable source at answer time — is the standard engineering mitigation, with the caveat that it makes the retrieval quality your new weakest link.
  • Evaluation against ground truth with human oversight, documented fact-checking procedures and logging of reported errors are the controls NIST AI 600-1 recommends for confabulation specifically.
  • Scoring that punishes confident error more than uncertainty, per the OpenAI paper, applied both in training and in how you grade your own system’s outputs.

Frequently asked questions

Is an AI hallucination the same as lying? No. Lying needs intent and knowledge of the truth; a model has neither. It generates the most plausible continuation, and a false continuation can be the most plausible one. The practical risk to you is the same, which is why the distinction does not make the output safer — only easier to explain.

Do newer models hallucinate less? Generally yes on common benchmarks, and the improvement is real. But rates are not zero on any current model, and the failures that remain concentrate in exactly the places users check least: obscure facts, fresh events and invented references. “Less often” is not “safe to skip checking.” Which model you are actually talking to also depends on your plan — the tiering that sets things like ChatGPT’s file upload limits decides model access too, so two people asking the same question on different tiers can get measurably different reliability.

What is the difference between a hallucination and a normal error? A normal error is a wrong or outdated answer. A hallucination is an invented one — content with no basis in any source, presented with full confidence. A wrong date copied from bad data is an error; a complete fake citation is a hallucination.

Does asking the AI to cite sources fix it? It helps only if you open the citations. Models can and do invent references that look completely real, so an unchecked citation list can make a hallucination more convincing, not less.

Can hallucinations be completely eliminated? On current evidence, no. The prediction mechanism that makes these models useful is the same mechanism that produces confident falsehoods at the edges of their knowledge. The realistic goal — for users and builders alike — is fewer hallucinations, in fewer high-stakes places, with verification wrapped around the rest.

This article is a general explainer based on the published sources linked above — IBM’s research summary, the OpenAI researchers’ 2025 paper and NIST’s Generative AI Profile. It describes how these systems behave in general; it is not an assessment of any one product’s current accuracy.