A hallucination is a plausible answer that happens to be false
Ask a language model for the page number of a quote in a book it has read about but never seen, and it will give you a page number. It won't say "I cannot know that". The number is formatted correctly, sits in a sentence that sounds like every other citation in its training data, and is made up.
That is a hallucination: a fluent, confident statement with no fact behind it. The word is a metaphor and a slightly misleading one, because nothing is malfunctioning. The model was asked for the most likely continuation of a citation and it produced one. On a question where the most likely-sounding answer is also the true one, which is most questions, the same process is what makes the model useful.
The forms it takes are predictable: invented references and URLs, statistics with no source, quotes attributed to real people who never said them, features of a product that don't exist, and legal or medical claims delivered as settled fact. Why AI gets things wrong walks through catching each type; this page is about why they happen at all.
Why a model that read the whole internet still makes things up
Two mechanisms, and they stack.
The model predicts, it does not retrieve. A language model has no index of facts. Everything it read was compressed into its parameters, and at answer time it regenerates text one token at a time from that compression, picking each token by probability. When the true answer was common in training data, the probable continuation is the true one. When it was rare, absent, or contradicted, the model still has to pick a token, and it picks a plausible one. What is an LLM explains the architecture; the short version is that there is no step where the model asks itself whether the sentence it is writing is real.
Training rewards guessing. In September 2025, OpenAI published a paper titled "Why Language Models Hallucinate" arguing that the benchmarks used to grade models score a wrong answer and a blank answer the same way, zero, while a lucky guess scores full marks. A model optimised against that scoring learns the same lesson a student learns on a multiple-choice test with no penalty for wrong answers: always guess. The paper's proposal is to change the grading so that "I do not know" is worth more than a confident miss.
Two court cases that put a dollar figure on a hallucination
Mata v. Avianca, New York, 2023. Two lawyers filed a brief citing six court decisions that ChatGPT had produced. None of the six existed. The judge sanctioned the lawyers $5,000 in June 2023 and made them write to the judges the invented opinions had been attributed to. Screenshots filed with the lawyer's affidavit show him asking ChatGPT whether the cases were real, and ChatGPT confirming that they were.
Moffatt v. Air Canada, British Columbia, 2024. Air Canada's website chatbot told a customer he could book a full-fare ticket for a funeral and claim a bereavement discount afterwards. The airline's actual policy said the opposite. When Air Canada argued the chatbot was a separate legal entity responsible for its own words, the Civil Resolution Tribunal rejected that in February 2024 and ordered the airline to pay C$812.02.
Neither model was broken. Each produced the most plausible-sounding answer to a question whose true answer wasn't in its reach. The cost landed on the person who did not check.
Hallucination is one of four failure types, and the fixes differ
People use the word for every wrong answer, and that hides the fact that the four common failures have four different causes.
| Failure | What happened | How to spot it | The fix |
|---|---|---|---|
| Hallucination | The model invented a fact to complete a plausible sentence | Specific, checkable, unsourced | Ask for the source, then open it; ask a second model |
| Stale knowledge | The fact was true when training ended and is not now | Prices, versions, office holders, anything dated | Turn on web search or paste the current page |
| Retrieval miss | Search or the uploaded file returned the wrong passage | The answer cites your document but addresses a different question | Read the cited passage, not just the citation |
| Misreading | The model answered a slightly different question than you asked | The answer is right about something else | Restate the question with the constraint made explicit |
Only the first row is a hallucination in the strict sense. Rows two and three are grounding problems and are largely solved by giving the model something to read. Row four is a prompt problem. Treating all four as one mysterious flaw leads people to the wrong fix, usually "use a bigger model", which helps least with row one.
The two-model check, and what it costs
The most reliable habit for the strict kind of hallucination is to ask a second, different model the same question and compare. Two models from different labs, trained on different data with different post-training, rarely invent the same citation. Agreement is weak evidence that the fact is real. Disagreement is a strong signal to go and check before acting.
The reason this works where "ask the same model to double-check" doesn't: a second pass by the same model is the same predictor guessing again from the same compression. A different model is a different compression.
The second opinion does not have to come from an expensive model. On the Artificial Analysis Intelligence Index, a composite of public reasoning, coding and knowledge benchmarks read on 2026-09-17, DeepSeek V4 Flash scores 40 and costs 1 Whizi credit a message; Claude Sonnet 5 scores 38 and costs 10. The dotted line is the frontier, the best score at each price, and every point on it is a different lab. A cross-check from two labs at 1 and 10 credits is 11 credits, and the side-by-side view sends one question to both on one screen. The lawyers in Mata v. Avianca paid $5,000 for the check they did not run.
What reduces hallucinations, ranked by how much
- Grounding. Web search, an uploaded document, or a pasted source turns the task from "recall" into "read and summarise", which models do far more reliably. This is the single largest reduction and it is a product setting, not a prompt trick.
- A second model. Different training, different failure modes. Catches most fabricated specifics.
- Asking for sources and opening them. A model asked to cite will cite; the check is whether the citation exists and says what the answer claims.
- Permission to say no. "If you are not sure, say so" changes what the most likely continuation is, so on hard questions the model declines more often instead of inventing.
- A bigger model. Frontier models hallucinate less on common knowledge than small ones, but the gap is smallest exactly where it matters, on rare and specific facts.
The order is the point. Most people reach for item five first, and it is the weakest lever on the list.
- Treat any specific, checkable claim with no source as unverified: a citation, a statistic, a quote, a product feature
- Turn on web search or attach the document whenever the answer depends on anything recent, private or precise
- Ask a second model from a different lab before acting on a number or a citation; disagreement means check
- Open the cited source and confirm it says what the answer claims, not just that it exists
- Add "say so if you are not sure" to prompts on hard questions
Frequently asked questions
Why do AI models hallucinate?
Because they generate the most plausible next words rather than look facts up, and because their training rewards a confident guess over an admission of ignorance. OpenAI's September 2025 paper "Why Language Models Hallucinate" shows the benchmarks score a blank answer and a wrong answer identically, so guessing is the winning strategy the model learns.
Can AI hallucinations be completely prevented?
Not with current models, because generating plausible text is the same mechanism that produces both correct answers and invented ones. They can be cut substantially: grounding the model in search results or a document, asking a second model, and demanding sources you then open each remove a large share. What remains is the residual risk on rare, specific facts.
Which AI hallucinates the least?
Frontier models from OpenAI, Anthropic and Google hallucinate less than small models on common knowledge, but the difference narrows on rare and specific facts, which is where the damage is. A grounded small model with web search on typically beats an ungrounded frontier model on factual questions, so the setting matters more than the badge.
Is a hallucination the same as a mistake?
It is one kind. A hallucination is an invented fact presented confidently. Stale knowledge (true at training time, false now), a retrieval miss (the wrong passage was fetched) and a misread question are different failures with different fixes, and only the first responds to "ask for sources".
What happened in the ChatGPT lawyer case?
In Mata v. Avianca, filed in New York, two lawyers submitted a brief citing six cases ChatGPT had invented. The court sanctioned them $5,000 in June 2023. A year later, in Moffatt v. Air Canada, a tribunal ordered the airline to pay C$812.02 after its chatbot invented a bereavement refund policy, rejecting the argument that the chatbot was responsible for its own statements.
