How to stop checking everything AI writes: what to check, what to skip

The rule in one line: check what must be true, skip what only must read well

The table above is the whole rule. The top five rows are your red flags. The bottom three are your skip list.

Here is why it works. AI does not make mistakes at random. It makes the same kinds of mistakes again and again, and they all sit in one place: facts. Anything that has to be true, not just sound right, goes on the red-flag side.

Researchers at Harvard and BCG watched 758 consultants use AI. On tasks the tool was good at, they did about 12% more work in about a quarter less time. On a task just past the tool's edge, the AI users were 19% less likely to get it right than people working without it. The study, in Organization Science, called this the jagged frontier. The tool stays just as sure of itself on both sides of the line.

That is the trap. The draft never warns you.

Every check goes against something that is not the draft

This is the part most people get wrong. They read the price inside the email to make sure the price is right. That proves nothing. The email is where the mistake lives.

So each red flag needs a home. A real place you look. "Check the price" is a wish. "Check the price against the current package sheet" is a check. Under pressure, the wish loses to "looks fine, send." The check with a named source gets done.

I think this is the most useful single habit in the whole book, because it turns a fuzzy worry into something you can finish.

Find your own red flags in your first ten runs

Don't copy the list above and call it done. Use it as a start.

For the first ten or so times a task runs, read the full draft. Each time something is wrong, write down the kind of mistake, not the mistake itself. Not "it said $1,800." Write "made up a price." Not "wrong wedding date." Write "got a date wrong."

Within a week or two the pattern shows. The same few kinds keep coming back. Other things are fine every single time. The ones that keep coming back are your red flags. The ones that are always fine are your skip list.

Then keep the red-flag list short. Three to five groups. If it runs to fifteen, you will skip it on the first busy day.

Different tasks fail in different places

Many anti-mistake checklists online give one list for every job. The book warns against that. Each task has its own weak spot, and your first ten runs show you where it is.

Two examples from the chapter. A routine that drafts social captions tends to make up hashtags and claim too much. A routine that sums up a contract tends to quietly drop a clause that changes the meaning. A client email fails on prices and dates. Same tool, three different red-flag lists.

So keep one short gate per task, not one for everything. When you set up a new routine, start a fresh log for it and let its own runs tell you what to check.

What if the problem is how it sounds, not what it says?

A lot of the "check AI writing" advice out there is about style: stock phrases, the same sentence shapes, a voice that sounds like nobody. That is a real problem. It is just not one you fix by reading every draft.

Fix it upstream, in the brief. Every phrase you delete twice goes on the brief's don't-list once, like "I hope this finds you well" or "excited." After that, the draft stops making the mistake, and tone can stay on your skip list. See why ChatGPT emails sound generic for the full brief.

The day the draft quoted the wrong package

Nadia, the wedding photographer in the book, had her reply drafts running well. They sounded like her. For a while she barely read them.

Then one came back smooth and wrong. A couple asked about one package, and the draft quoted the price of a different one, off by several hundred dollars. She nearly sent it. She only caught it because the number looked a little low.

It sounded so sure of itself.

She did not go back to reading every line. A full read had already missed it. She built the two lists, taped the red flags to the top of the drafts, and her check went from a guilty two-minute skim to a thirty-second pass. Price against the sheet. Date against the calendar. Names against the couple's email.

A month later the same list caught a draft that offered a Saturday she had already booked. The tool had no idea what her calendar said. It just gave the friendliest answer.

Skipping on purpose is part of the rule

The skip list is a promise, not a hint. If you write "tone" on it and then read for tone anyway, you have not built a check. You have built a list you ignore.

It takes a little nerve the first few times. But the choice was made on evidence, a dozen clean runs, not on a whim. Trust what you collected.

And when one red flag costs a lot, like a price a client can hold you to, go one step further. Leave a blank where the price goes so the draft can never fill it.

Where this sits in The Leverage Stack

This is the problem The Leverage Stack was written for, and it gets a full chapter.

Chapter 5, The Quality Gate, walks you through the log, the two lists, and how to weigh each red flag by what a mistake would cost you. It ends with the Quality Gate Checklist, one page you fill once per task.

Chapter 3, The Brief, cuts the mistakes before they start. Its don't-list is where every fix you make twice gets written down once.

The book leans on two studies, and you can read both yourself: the jagged frontier paper above and the NBER study of 5,179 support agents, where the least experienced workers gained the most from AI help. The checklist is also in the free workbook. If the checking feels like a symptom of a bigger jam, start with why a one-person business can't keep up.