The dangerous thing about a modern AI model is not that it gets things wrong. It is that the wrong answers look exactly like the right ones — same tone, same structure, same quiet confidence. There is no wobble in the voice when it invents a statistic.
I use AI every day for teaching, writing and client work, and I have never once published something it wrote without checking it. Here is the five-minute routine I run instead, and the specific places I have learned to look first.
Know Where the Errors Actually Live
You cannot check everything, so check the things that break. In practice, AI output goes wrong in four predictable places: numbers, names, dates and citations. Everything else — explanations, structure, reasoning, tone — is usually fine or obviously off.
So the first move is mechanical. Read the draft once and mark every figure, proper noun, date and source. On a 900-word article that is typically eight to twelve items. Those are your five minutes. The prose between them is not where the risk is.
Make the Model Mark Its Own Confidence
Before you check anything manually, make the AI narrow the search for you. It has a rough sense of which claims are load-bearing guesses, but it will never volunteer that unless you ask.
This is not a lie detector, and a confident model can be confidently wrong. But it is a very good triage tool: the "low" and "medium" rows are almost always where the real problems sit, and it turns a vague worry into a short, checkable list.
Verify Sources by Opening Them
The single most common AI failure I still see in 2026 is the plausible citation — a real journal, a real-sounding author, a paper that does not exist. Or worse, a paper that does exist but says something different.
The rule is simple and non-negotiable: if you have not opened the source, you have not checked the claim. Not searched the title, not seen it referenced — opened it and found the sentence. A source that takes more than sixty seconds to locate should be cut, not chased.
The same applies to statistics. "Roughly 70% of small businesses..." with no origin is not a fact, it is a texture. Cut it or source it.
Run the Second-Model Check
For anything going out publicly, I paste the draft into a different model and ask it to attack the piece rather than praise it. Different training, different failure modes — one model's blind spot is often another's obvious catch.
The "do not improve the writing" line matters. Left unconstrained, any model will happily rewrite your voice into beige and call it feedback. You want a fact-checker here, not a co-author.
Apply the Domain Test
The last check is the one no tool can do for you: does this match what you know from experience? Every time I catch a genuinely subtle AI error in physics, it is not because I looked it up. It is because the explanation was technically defensible but pedagogically wrong — the kind of thing that sounds fine until a student tries to use it.
Which means AI is safest exactly where you are strongest, and most dangerous where you are least able to spot the flaw. If you are publishing outside your expertise, either get a human who has it to read the piece, or do not publish it. That is the same standard I hold my own work to when I build websites and content systems for clients — the AI drafts, a human signs off.
The Five Minutes, In Order
Mark the numbers, names, dates and citations. Ask the model to rate its own confidence in each. Open every source or delete it. Run the draft past a second model with an adversarial prompt. Then read it once yourself against what you actually know.
That is it. It costs you five minutes on a piece that took an hour, and it is the difference between AI making you faster and AI making you faster at being wrong in public.