The 60-second verification pass

Speed without verification isn't productivity. It's exposure.

The problem with most advice on checking AI output is that it's too thorough to survive contact with a real working day. A twelve-point checklist gets used twice and then abandoned. What you need is something short enough that it becomes automatic.

There is a reason the budget is sixty seconds and not five minutes. A check you do every time beats a better check you do when you remember, and what you can afford every time, on every draft, is about a minute. So the pass below is not everything you could check. It is the four places where fabrication actually lives, which is what makes a minute enough.

Here is that pass. Four steps, fifteen seconds each.

1. Numbers

Highlight every digit on screen. For each one, ask: did I supply this, or did the model?

Anything the model supplied gets checked at source or gets deleted. No exceptions, no "it's probably right." Approximately-right numbers are worse than obviously wrong ones, because nobody catches them.

Watch it work on a specimen. You paste your notes into the tool, "renewal is $18,400, up from $17,100", and ask for a client summary. The draft comes back: "The renewal price is $18,400, an increase of roughly 7.6% on last year." You supplied both prices. The model supplied the percentage, and this time it happens to be right. Next week, same task, the percentage will be confidently wrong, because arithmetic is exactly the kind of work these tools get wrong while sounding right, and a plausible-but-wrong 7.6% survives every glance-level read. That is why the rule is mechanical rather than judgement-based: model-supplied numbers get recomputed or removed. Even the ones that look right. Especially those.

2. Names and citations

Every person, company, product, paper, law and URL. Anything you cannot personally vouch for gets searched.

If the output includes links, open one at random. If it doesn't check out, open all of them. A cited link is not a verified claim, a model can link a real page and misrepresent what it says, which passes a glance-level check and fails a real one.

The dangerous case is not the invented source, which a search catches in seconds. It is the real source, correctly titled, correctly linked, that says something adjacent to what the draft claims it says. "The regulator's 2025 guidance permits this" over a working link to the regulator's 2025 guidance takes the model eleven words to produce, and takes you one click and three minutes of reading to actually verify. The click and the reading are the check. Open the page, find the passage, read it as a sceptic, and if you cannot find the sentence that supports the claim, the claim comes out, however good the link looks.

3. The confident middle

Fabrication rarely sits in the opening or the closing. Those are formulaic and the model is on solid ground there.

It sits in the specific, confident, load-bearing sentence in the middle of paragraph three. Find the sentence a reader is most likely to act on. Check that one hardest.

There is a reason for the geography. Openings and closings are the most template-shaped parts of any document, and templates are what these tools do best. The middle is where the specifics live, and specifics are what a plausibility engine invents. So skim the pleasantries and hunt for the sentence that changes what the reader does next: the deadline, the commitment, the "the contract allows this", the "the team has capacity for it". If a reader would forward it, quote it, or act on it, that sentence gets the hard check.

4. Absence

What's missing?

The model can only work with what you gave it. The caveat you know about, the stakeholder you forgot to mention, the risk that didn't make it into your notes, none of that appears, and nothing flags its absence.

The model's errors of omission become yours, silently. This is the step people skip and the one that causes real damage.

The working version of the check is one question: what would a careful reader expect to find here that isn't here? A renewal summary that never mentions the notice period. Meeting notes that record every agreement and omit the one disagreement. The first three steps are about what is on the screen; this one is about what isn't, and it is the step no tool can do for you, because only you know what you know and the model doesn't.

Scale it to the stakes

An internal Slack summary doesn't need this. Anything going to a client, a regulator, a board or the public gets the full pass plus a second human.

The scaling rule in one line: the cost of the check should track the cost of being wrong. A message you could correct thirty seconds later with "sorry, typo" needs nothing. A number that will be pasted into someone's board pack will surface wrong in front of exactly the audience you least want, weeks after you could have quietly fixed it.

Attach it to something you already do

A pass this short survives only if it has a trigger, and the trigger is the send button. Not a scheduled "verification time", which stops happening within a week, but a rule about the cursor: it does not reach send until the digits have been highlighted. People who bind the habit to an action they already perform are still doing it a year later, which is the entire point of making it sixty seconds in the first place.

The rule underneath all of it

Your name is on it. Not the model's.

"The AI wrote it" has never once been an acceptable answer to "why is this wrong?", and it never will be. You are the author of everything you send, regardless of how it was produced, and the sixty seconds is what makes that a comfortable position rather than a risky one.