AI hallucinations. When a model makes something up, and says it with the same confidence as the truth.

A hallucination is a false statement produced by an artificial intelligence model with the same fluency and confidence as a true one: a figure, a quotation, a rule, a price that do not exist. The risk lies in the confidence with which it is stated.

This entry is part of the glossary of AI and automation, where the term appears as hallucination. Here the definition widens: what they are, why a model makes things up, how often it happens according to published measurements, who is liable, and what makes them rare.

What hallucinations are

In an article of 5 September 2025 OpenAI defines hallucinations as «plausible but false statements generated by language models». They can concern anything: a date, a name, a rule, the title of a book, a number in a document the model was only supposed to read.

The example used by the authors themselves is telling. Asked for the title of the doctoral thesis of one of them, a widely used chatbot confidently gave three different answers, all wrong; asked for his date of birth, three different dates, also all wrong.

Why a model makes things up

An LLM learns to predict the next word on huge amounts of text, and in that text no label says what is true and what is false. Patterns, such as spelling, are learned well. Rare, arbitrary facts, such as the date of birth of a little-known person, cannot be inferred from any pattern, and there the model guesses.

The second cause lies in evaluations. According to OpenAI most tests measure only right answers, and saying «I don't know» scores zero: as in a multiple-choice quiz, guessing pays. A model trained to do well in those tests learns to always answer, even when it does not know.

How often they are wrong: the published measurements

Published measurements vary a lot with the task and with the possibility of abstaining, and none is a forecast for your case. Taken together, they show where the risk lies: in the questions the model cannot answer and answers anyway.

Measurement	What it found	On which task
OpenAISeptember 2025	A model that abstains on 52% of questions gets 26% wrong; one that abstains on 1% gets 75% wrong.	Questions with a single right answer, from the SimpleQA test.
Stanford RegLab2024	Three legal research tools sold as free of hallucinations are wrong between 17% and 33% of the time.	Questions of US law, with documents available to the system.
Itriatest bench of 4 September 2026	An unreadable PDF, placed on purpose in a batch of 30 documents, set aside as «other» without an invented field.	Data extraction from orders, delivery notes and invoices.

The first row holds the most useful lesson. The two OpenAI models give almost the same share of right answers, 22% against 24%, but the first gets 26 in a hundred wrong instead of 75, because when it does not know it abstains. In a business an «I don't know» costs a question passed to a colleague; an invented answer can cost a customer.

Who is liable: a case decided in court

On 14 February 2024 the Civil Resolution Tribunal of British Columbia, in Canada, decided the case Moffatt v. Air Canada. The chatbot on the airline's website had given a customer wrong information about a bereavement fare, and the airline argued it could not be held liable for what the chatbot said.

The tribunal rejected the argument, in paragraph 27: «It should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot». The airline was ordered to refund the fare difference. What your assistant says, for the customer, you say.

Can they be avoided?

They can be made rare, and the key is abstention. OpenAI puts it this way: accuracy will never reach 100%, because some real-world questions have no answer that can be worked out, but hallucinations are not inevitable, because a model can abstain when it is not sure. In a business system three precautions are needed.

The right sources at the moment of the question. A system that answers from the company's documents, with the technique called RAG, can say which document each answer comes from, and whoever reads it can check it.

Permission to say «I don't know». The system has to be built so that, when a source is missing, it stops and passes the question to a person, instead of completing the answer anyway.

A person where errors are costly. Prices, conditions, deadlines and everything that commits the company go through a human in the loop before sending.

How Itria handles them

In Itria's systems the rule is written before the code: when proof is missing, the system stops and alerts a person. On the test bench of 4 September 2026 the unreadable document placed on purpose was set aside as «other», without a single invented field, and across the whole batch 234 of 237 extracted fields were correct.

The knowledge-base chatbot answers from the documents approved by the company and states that it is a system. The principles are written in Ethics, and here the third matters most: the machine prepares, but a person always decides.

Related terms

LLM

The language model that produces the text, and with the text the hallucinations. Understanding how it works explains why it makes things up.

Human in the loop

The person who approves before sending. It is the last line of defence, and it belongs where errors are costly.

RAG

The technique that gives the model the right documents at the moment of the question, so that it answers from them.

Knowledge base

The set of approved documents a system can answer from. If it is out of date, answers are wrong even without hallucinations.

Questions and answers