Guide
How to check whether an AI answer is right
AI tools don't look things up the way a search engine does. They predict a likely answer. Most of the time that works; when it doesn't, the answer still sounds right. Here is how to catch it.
Where AI answers fail most
Be most careful with:
- specific names, dates, figures and quotations
- citations, case references and links, which can be invented
- anything recent, niche or local, where there was little to learn from
- questions that contain a wrong assumption, which the tool will often go along with
Five checks
- Specifics: underline every name, number and date, and check each one.
- Sources: ask where each claim came from, then open the source yourself.
- The premise: ask whether your question assumed something untrue.
- The counter-case: ask for the strongest argument against the answer.
- A second opinion: ask a different tool, or a person who knows, for anything that matters.
Match the effort to the stakes
A rewording of your own email needs a read-through. A figure going into a board paper needs all five checks and a note of where it came from. Deciding the stakes first saves time on the low-risk work and catches the errors that matter.