You're not going to fact-check every text you get AI to help with. Nobody does. But some things need it: a client email, a social post, an invoice, anything you're about to publish or send.
Here's a five-step routine. Run through it in under two minutes.
Step 1: Spot the claims
Read the AI's answer and underline (or highlight) anything that's a specific fact. Names of people. Dates. Numbers. Quotes. Citations. Product features. Legal claims.
Ignore opinions and general advice. Focus on the checkable bits.
If there are no specific claims, you can usually skip to step 5.
Step 2: Ask "would I stake £50 on this being right?"
For each highlighted claim, pause and ask yourself:
If I was betting fifty quid on this being exactly right, would I still say yes?
This is a quick gut check. If you'd hesitate, it needs verifying. If it's a name you've never heard, a number that feels round, or a date that's suspicious, it needs verifying.
Step 3: Verify with a real source
Open a new tab. Don't ask the AI to check its own answer (it will often "confirm" the same hallucination). Do one of these instead.
- Google it. Especially for facts about real-world people, places, and events.
- Check the actual source. If the AI quoted a report, find the report. If it quoted a law, find the Act.
- Ask a different AI with search turned on. Gemini (with web) and ChatGPT (with browsing) will often catch mistakes that the first AI made.
If the fact doesn't show up in a real source in under 30 seconds, treat it as made up and cut it or rewrite without it.
Step 4: Check the numbers
Numbers deserve their own check because they're the easiest things to get away with.
- Does the maths work? If the AI says "a 30% increase from £40 to £60", do the maths. (That's 50%, by the way.)
- Is the date plausible? Anything in "last year" or "recently" should be double-checked against today's actual date.
- Is the unit right? The AI sometimes swaps pounds and dollars, millions and billions.
Step 5: Read it as a sceptical human
Last step. Put the AI out of your mind and read the whole thing as someone who doesn't know you used AI. Three questions:
- Does it sound like me? If it doesn't, edit it until it does.
- Would I send this if I'd written it myself? If not, something's off.
- Is there anything here I'd be embarrassed to defend? Cut it.
That's it. Two minutes, five steps. Saves you from the awkward apology email later.
The meta-rule
Trust AI to help you think. Don't trust it to be right without checking.
Print that on a sticker for the side of your laptop.
Next up: Module 2 kicks off with the anatomy of a good prompt. We move from understanding AI to writing prompts that get great answers first time.