What is an AI hallucination? Definition and examples

An AI hallucination is a fluent, confident answer from a language model that is false or unsupported, such as an invented price, policy or citation.

How it works

A large language model writes by predicting likely next words. It has no built-in sense of "I don't know", so when it lacks a fact it can still produce text that sounds right. A September 2025 research paper, Why Language Models Hallucinate, argues that models hallucinate because "the training and evaluation procedures reward guessing over acknowledging uncertainty": most benchmarks score a confident guess higher than an honest "I'm not sure" (Kalai et al.).

Hallucinations get more likely when:

• the question is about something specific to your company that the model never saw;
• the retrieved sources are missing, outdated or contradict each other;
• the question invites a yes or no answer the sources don't support ("Do you integrate with X?").

Why it matters for your website

On a company website a hallucination is not a quirk; it is a statement made by your business. In Moffatt v. Air Canada (February 14, 2024), British Columbia's Civil Resolution Tribunal dealt with an airline chatbot that told a customer they could apply for a bereavement fare refund after traveling, which contradicted the airline's actual policy. The tribunal called the suggestion that the chatbot was "a separate legal entity that is responsible for its own actions" a "remarkable submission", held that the airline "is responsible for all the information on its website", whether it comes "from a static page or a chatbot", and ordered it to pay CA$812.02 in damages, interest and fees (decision; summary on Wikipedia).

The practical defenses are well known:

• Ground answers in your own pages with retrieval-augmented generation and show the source link.
• Allow "I don't know." Instruct the assistant to say when the sources don't cover a question, and make that a good outcome, not a failure.
• Route sensitive topics such as refunds, contracts and medical or legal questions to a person through human handoff.
• Review real conversations regularly and fix the missing or unclear pages behind wrong answers.

Example

A visitor asks: "Does the Pro plan include phone support?"

• Hallucinated: "Yes, Pro includes 24/7 phone support with a dedicated manager." Nothing on the site says this.
• Grounded: "Our pricing page lists email support for Pro and doesn't mention phone support. I can ask the team to confirm. What's the best email to reach you?"

Related terms

• Grounding: tying answers to verifiable sources.
• Retrieval-augmented generation (RAG): fetching your content before answering.
• Large language model (LLM): the kind of model that hallucinates.
• Human handoff: passing risky questions to a person.