Check the answer before you use it
AI answers sound confident whether they are right or wrong. A fluent paragraph is not evidence. Treat the first answer as a draft from a smart colleague who sometimes makes things up: useful, and worth checking.
What to check
- Facts, names and dates. Look up anything you will repeat to someone else.
- Numbers and math. Recalculate totals and percentages yourself or in a spreadsheet.
- Quotes and sources. If it cites a book, a study or a link, open it. Make sure it exists and says what the AI claims.
- Code. Run it, and run its tests.
- Recent events. An AI may not know about something that happened after it was trained, unless it can search the web.
The more it matters, the more you check. A lunch suggestion needs no checking. A figure in a client report does.
Ask it to help you check
You can use the AI as part of the check:
- "What assumptions did you make?"
- "Which parts of this answer are you least sure about?"
- "Give the sources, and tell me if you cannot find one."
- "Show your steps so I can follow them."
These questions will not make an answer correct, but they show you where to look.
"Are you sure?" is not a test
Asking "Are you sure?" often makes the AI change a correct answer, or defend a wrong one. It is a poor test. Better to ask for the reasoning, or to look at the claim yourself.
Get a second look
Try the question another way, or in a new chat, and compare. If two attempts disagree, you have found something to verify. Comparing answers from two different models is another good check.
High-stakes topics
For health, legal, tax and money decisions, use the AI to prepare: to explain terms, to draft questions, to organize your notes. Then take the questions to a qualified person. The AI is not a substitute for one.
Keep a check habit
A quick routine works well:
- Mark every claim you will rely on.
- Verify those, not everything.
- Ask the AI to fix what you found wrong, with the correct fact.
Try it
Ask an AI for five facts about a topic you know well. Count how many are right. That is your own calibration for how much to trust it on topics you do not know.