The more I used AI, the more I trusted it. Somewhere along the way I stopped double-checking most of what it gave me.
Then it handed me a number for a client email. A clean stat, a study name, a year. It looked exactly right, so I almost sent it.
Something made me check first. The study didn't exist. The model had invented the name, the year, all of it. A made-up fact, seconds from going out under my name. After that I read everything twice. Sometimes three times, until I wasn't sure AI was saving me any time at all.
You told me the same thing. One of you caught it quoting a court case that never existed. Another said it keeps changing what they actually wrote, even after telling it not to. Someone almost paid an employee for two days he never worked, because it invented the hours and they nearly sent it without checking.
If you've been rereading every answer since your own version of that, that makes sense. You got burned once, so now you check everything. And that checking is what eats the time AI was supposed to save.
You don't have to check every line. The same few things are usually where mistakes show up.
Where it invents (check these every time)
It makes things up when it's reaching for something it doesn't actually have in front of it:
Specific numbers and statistics
Sources, citations, links, study names
Dates, counts, and any "how many"
Quotes, and who said them
Recent events, prices, anything that changed lately
Legal, medical, or financial specifics
Where it's steady (you can usually relax)
It holds up when it's working with what you already gave it, or shaping words instead of fetching facts:
Rephrasing or tightening your own text
Outlining and structuring
Brainstorming options for you to choose from
Summarizing a document you pasted in
Changing tone, or explaining a concept you can sanity-check yourself
🌋 The pressure test (one prompt, ten seconds)
Before you paste an AI answer into anything that matters, run this on it:
Pressure-test your last answer before I use it.
Do all four, briefly:
1. Confidence: rate how sure you are, 1-10,
and say what would lower it.
2. Weak points: flag every specific number,
date, name, quote, or source
you gave.
For each, say "verified from my knowledge" or "you should
check this."
3. Assumption check: what did you assume about my situation that,
if wrong,
breaks the answer?
4. Verify steps: tell me the fastest way to confirm the flagged items,
what to search, what to look at.
Don't defend the answer. Just show me where it's soft.
You can copy it once. Better: turn it into a shortcut called ;check, so you do not have to come back and find it.
I showed the setup here: [turn your best prompts into shortcuts].
Run it on the very next answer, not "next time." Ten seconds today is how you stop rereading everything twice.
You don’t need to remember to check. You need a shortcut that makes checking easier than skipping it.
⏱️ Your 10-second win
The next time ChatGPT writes something you're about to send, run the pressure test first. Check the points it flags, then send it.
Run it on the very next answer, not "next time." Ten seconds today is how you stop rereading everything twice.
How was today's issue?
One honest cavet
The model is still checking its own work, so treat the result as a second look, not a final answer. If it says something looks correct, that's a reason to feel more confident, not a reason to stop checking. Anything important, especially numbers, quotes, or information going to a client, still deserves a quick check against a real source.
One more thing about confidence scores. If the model says it's "8 out of 10" confident, don't treat that like a real measurement. Think of it as a rough signal that helps you decide what to check first. It's not proof that something is correct.
You'll also see prompts online that say things like, "Only answer if your confidence is above 0.8." Skip those. The model can always give itself a high score, even when it's wrong. The safer approach is simple: verify the answer, not the confidence score.
If you want to actually use this check instead of saving it and forgetting it, turn the pressure-test prompt into a shortcut called ;check. I showed the 2-minute setup here: [turn your best prompts into shortcuts].
Next week I'm opening a room where people compare what it caught and what it still let slip, so you pick up the failure patterns faster than checking alone ever touches them.
Dan Rice · AI Signal Read once. Use AI better all week.