How do I catch AI mistakes and hallucinations?

AI assistants can state wrong information with total confidence. This is called a hallucination, and a made-up number looks exactly like a real one. The book's answer is the producer-to-editor mindset: the AI produces the draft and you edit it before anything goes out. Chapter 8, When Claude Gets Confused, is dedicated to this topic.

Check before you send

Ask three questions about any output that matters:

1. Does it match what I already know? If the AI says the project is on track and you've heard about delays all week, stop and check.
2. Can I verify it against a source? Compare every name, number, date, and commitment with your documents. Most checks take seconds.
3. Does it pass the common-sense test? A plan that looks too simple for a hard problem usually is.

Match the review to the stakes

• Low stakes (brainstorming, first drafts): a quick read for reasonableness.
• Medium stakes (internal updates): spot-check a few specific claims.
• High stakes (executive reports, budgets, commitments to clients): verify every fact.

Reduce errors at the source

• Give it the facts. Most mistakes come from missing context. Keep your core documents in the assistant's workspace. See What documents should I give my AI assistant?.
• Tell it not to guess. Ask the assistant to say "not in the documents" when it can't find something, and to point to the document an answer came from.
• Remove conflicting files. Two versions of the same schedule lead to blended, wrong answers.
• Restart drifting conversations. Long chats can mix up details, so start fresh with a summary.

When you find an error, work out why it happened (missing context, a guess, or a misunderstood instruction) and fix that, not just the one output.

For more, read AI Makes Up Facts: How to Detect and Stop Hallucinations, How to Know When AI Is Wrong: The Three-Question Test, and AI Mixes Up Your Project Details.