What Is an AI Hallucination and How Do I Manage It?

An AI hallucination is an answer that sounds confident and is wrong, invented or unsupported by any source. You cannot remove it entirely, so you manage it with checks.

Why it happens

Language models predict plausible text. They do not look facts up in a reliable database unless connected to one, as with retrieval-augmented generation. When a model lacks the information, it can still produce a fluent answer, including made-up names, figures, citations or policies.

Typical examples

• A cited court case or article that does not exist.
• A product feature your supplier never offered.
• A wrong total in a summary of figures.
• A policy statement that contradicts your own terms.

How to manage it

1. Supply the source. Give the model your document or data and ask it to answer only from that, and to say when the answer is not there.
2. Ask for quotes or references from the source, then spot check them.
3. Verify facts that matter. Dates, amounts, names, legal and medical statements must be checked by a person.
4. Keep humans in the loop for anything customer-facing. Wikipedia's entry on AI hallucination lists more examples.
5. Use the right tool. Calculations belong in a spreadsheet, not in a chat reply.
6. Teach staff the habit. Treat AI output as a draft from a keen junior who sometimes bluffs.

Where it matters most

Legal, financial, health and compliance content, anything sent under your name, such as quotes and proposals, and anything that feeds a decision. For low-stakes drafting, the risk is small.

A short checklist next to the tool, covering names, numbers, claims and tone, makes review quick and routine.

21Labs builds review steps into the habits it sets up through AI adoption.