The dangerous mistake is the plausible one

When AI gets something obviously wrong, you catch it without trying. The real risk is the believable error: the supplier name that is almost right, the percentage that sounds reasonable, the policy clause that reads like every other clause. These slip through because nothing about them looks wrong. Newer models invent less than older ones did, but "less" is not "never", and the mistakes that survive are precisely the ones that do not announce themselves.

The fix is not to stop using AI. It is to run a short, consistent check before anything you produce with AI goes to another person. Once it becomes a habit it takes about two minutes.

The check has three passes

1. Facts and figures

Anything specific — a number, a name, a date, a financial or legal claim — gets traced back to your source document or a real website. If you cannot trace it to something solid, it does not go in. AI is very good at producing a number that looks right; it has no idea whether it is.

2. Quotes and sources

If the AI cites a study, a person, or a webpage, open it and confirm two things: that it exists, and that it actually says what the AI claims. Invented references that look entirely real are one of the most common AI errors going, and they are embarrassing precisely because they look credible.

3. The "would I stake my name on this" pass

Read the whole thing once as if you wrote every word — because to the person receiving it, you did. This last read catches tone problems and small claims the first two passes missed.

Let AI help you check AI

You can speed up the first pass by asking the tool to flag its own riskiest claims. It will not catch everything, but it surfaces the parts most worth your attention, so you are not re-reading every line with equal suspicion.

Below is a draft you produced for me. Before I use it, I need to check it. Go through it and list every specific factual claim — names, numbers, dates, statistics, quotes, and citations. For each one, mark it as: - [VERIFY] something I should confirm against a source - [SOURCE?] a claim that needs a reference you have not given - [SAFE] general wording that carries no factual risk Do not rewrite the draft. Just give me the checklist. Draft: [paste the AI output here]
💡 A faster version for high-stakes work

For anything important, ask the same question of a second AI tool — paste the claim into Claude or Gemini as well as ChatGPT and see whether they agree. When two different tools give the same answer it is more likely to be right; when they disagree, you have found exactly the thing to check by hand.

The mindset that keeps you safe

Treat every AI output as a strong first draft from a fast, confident assistant who is sometimes wrong and never tells you which times. That framing is not cynical, it is just accurate, and it keeps you checking the specifics rather than trusting the tone. The people who get caught out by AI are not the ones who use it most. They are the ones who stopped reading what it gave them.

One thing to do this week

The next time you use AI for something that goes to another person — an email, a report section, a summary for your manager — run the checklist prompt above before you send it. Count how many items come back marked [VERIFY] or [SOURCE?]. For most real work it is two or three, and checking them takes about two minutes. Do that once and you will see why it is worth making automatic.