AI Mistakes and How to Check Them

Introduction

AI can be wrong in a very convincing way. People call a made-up answer a hallucination. Learning to check AI is one of the most valuable skills you can build.

The simple explanation

AI predicts text that sounds right. Sounding right is not the same as being right.

Because a model produces likely text, it may fill a gap with something that looks believable, such as a book that does not exist, a wrong date or an invented quote. It does this without any warning, in the same confident tone it uses for correct answers.

Warning signs

Be extra careful when an answer contains:

• Specific numbers, dates, names or quotes
• Links, citations or article titles
• Medical, legal or money advice
• Very recent events
• Anything you plan to publish, submit or act on

A quick checking routine

1. Ask for sources, then open them yourself. Make sure they exist and say what the AI claims.
2. Check with a trusted source, such as official documentation, a textbook or a respected organisation.
3. Cross-check an important fact in at least two places.
4. Test it. If the answer is code or a calculation, run it and try examples.
5. Ask the AI to doubt itself, for example "Which parts of that answer are you least sure about?" It helps, but it is not proof.

Try it yourself

Ask a chatbot for three facts about a topic you know well. Mark each one true or false. How many were right? What kinds of detail were most likely to be wrong?

Summary

• A hallucination is a confident answer that is not true.
• Be most careful with numbers, quotes, links, advice and recent events.
• Verify important answers with trusted sources and by testing.