A note before the check. Tools change constantly, but the places they get things wrong have stayed stable across every model and every update. That stability is exactly what makes a fixed check worth having, whether you draft in ChatGPT, Claude or Gemini.
There is a particular hover that happens over the send button when AI helped write the thing underneath it. The report reads well and the email sounds right, but somewhere in the back of your mind is the knowledge that one detail in there might be wrong. So you either reread the whole piece with narrowed eyes, which defeats the point of the help, or you send it and carry a low hum of doubt into the afternoon.
There is a better option than either, and it rests on one useful fact about how these tools fail. The prose is almost never the problem. ChatGPT writes fluent, sensible sentences all day long, and the errors hide inside them in the same five places every time.
The short answer. AI's confident mistakes cluster in five predictable spots: numbers, names, dates, citations and quotes. Before anything AI helped write leaves your screen, scan just those five against what you know or can check. It takes under a minute for a normal email and a few minutes for a report, and it catches most of what could genuinely hurt you without proofreading everything twice.
Why these five
They share three properties that make them dangerous together. They are the load-bearing parts of any document, the bits a reader will act on, repeat or forward, so an error there travels. They are checkable, which means someone will eventually notice when one is wrong, and that someone is rarely you.
And they are exactly where the tool fills gaps by inventing, because a plausible figure or reference is easy to generate and hard to spot at a glance. The sentences around them stay innocent. The details do the lying.
The check itself
| Look at | What tends to go wrong | What to do |
|---|---|---|
| Numbers | Totals and percentages produced by prediction rather than calculation, or figures copied slightly wrong | Check every figure against the source, or redo the sum in a spreadsheet |
| Names | People, companies and products swapped for similar ones, or spellings drifting mid-document | Confirm each against the original, especially where two similar names exist |
| Dates | Deadlines shifted by a day or a year, or "next Friday" resolved to the wrong Friday | Check any date someone will act on against a calendar or the source |
| Citations | References that look perfect and do not exist, or exist and say something else | Open every link and confirm it says what the text claims |
| Quotes | Words attributed to people who said something adjacent, or nothing at all | Verify against the transcript, word for word if it will be repeated |
How hard to check, and when
The scan scales with where the writing is going. A Slack message to a colleague gets a glance at whichever of the five it contains. A figure heading into a board pack gets checked against the source before it goes anywhere, because a wrong number that gets forwarded is the most expensive kind of error.
For the fuller version of that judgement, how to tell when ChatGPT is making it up covers how hard to check based on stakes, and how to fact-check AI output covers the methods for verifying a specific claim. The pre-send check is the everyday habit that sits in front of both.
Make the tool do some of the work
One line added to any request shifts part of the burden back where it belongs. Ask it to flag anything it is not certain of and to say what you should double-check, and you will often get an honest map of the soft spots. Treat that map as a starting point rather than a guarantee, because the tool has no reliable sense of its own errors, but it narrows the search and costs nothing to ask.
The habit, in one sentence
Before anything AI touched leaves your screen, find the numbers, names, dates, citations and quotes, and look at each one for as long as it deserves. That single pass is the difference between using these tools with a low hum of anxiety and using them with your judgement in charge. It also takes less time than rereading the first paragraph suspiciously ever did.
Not sure where to start?
The right tool makes the check shorter.
The free finder tells you which AI to open for the task in front of you, in about 60 seconds, and the full "AI, sorted." reference lands in your inbox.
Try the free finder →One question in, one answer out.
Clair helps non-technical professionals know when to trust their AI, when to check it, and when to skip it.