What is AI hallucination?

An AI hallucination is when an AI tool states something false as if it were true, such as a made-up figure, quote, source or policy. It happens because language models write the most likely-sounding answer and do not check facts as they go. A false answer reads just as confidently as a true one. It is most common with numbers, names, dates, recent events and a company’s own information.

Why it matters for a business

A hallucinated answer looks exactly like a correct one, so it passes review at reading speed and gets acted on. Newer models and reasoning models hallucinate less, and no model is free of it. Two things limit the damage. Grounding puts the source document in front of the model before it answers, so it summarises rather than recalls. Checking the output before anyone relies on it catches what grounding misses. How much checking a task needs depends on what a wrong answer would cost.

Example

Let’s say someone asks a company assistant for last year’s Q3 sales. The model has never seen the company’s figures, so it returns a number in the right order of magnitude, formatted like every answer before it, and invented.

Go deeper

Why AI hallucinates (makes things up)

Related terms

• Grounding
• RAG (retrieval-augmented generation)
• Knowledge cutoff
• AI guardrails
• Training data

All terms, A to Z