How to Become the Person Who Checks the AI

Somebody has to sign it

Every organisation using AI seriously ends up with the same shape of problem. Output arrives quickly, it looks finished, and somebody still has to say yes before it goes to a customer, a regulator, a board or a court. That yes is a person. It cannot be anything else, because a yes is a promise, and only a person can make one.

That role is quietly growing everywhere, usually without a job title attached to it yet. It goes by review, or sign-off, or quality, or simply being the one people send things to before they send them out. Being genuinely good at it is one of the more durable positions you can hold.

The catch, and it is a real one, is that you cannot fake it.

Why checking is harder than it sounds

Reading through an AI output and nodding is not checking. It feels like checking, which is what makes it dangerous.

The reason is the fluency trap. Human writing gives off signals when the writer is unsure: hedging, awkwardness, thin detail. AI output gives off none of those. A wrong answer arrives in the same confident register as a right one, so your instincts, honed on humans, are calibrated for the wrong thing.

So the only reliable way to check is to know the answer yourself, or to know exactly where the answer can be verified. Which brings us back to depth.

The four places errors hide

Over time, checkers develop a mental list of where problems cluster. A reasonable starting list looks like this.

Specifics that sound retrievable. Names, dates, figures, citations, legal references, product specifications. Anything that looks looked-up is worth actually looking up.

Things that are true in general but not here. The single most common failure in professional contexts. The output describes standard practice, and your organisation, contract, jurisdiction or client is an exception. Nothing on the page is wrong in the abstract, and all of it is wrong for you.

What is missing. Absence is invisible. An output can be entirely accurate and still leave out the caveat, the stakeholder, the risk or the alternative that changes the decision. Reading for absence is a learned habit and almost nobody does it naturally.

Confident reasoning built on a wrong assumption. Check the premise before you check the logic, because sound reasoning from a false start produces a very persuasive wrong answer.

Below is a piece of work. Do not improve it and do not comment on the writing.
Give me four separate lists:

Every specific claim (names, dates, numbers, references, sources) that could be checked externally, with how to check each one.
Anything that is true as general practice but might not hold for a specific organisation, contract or jurisdiction.
What a knowledgeable reader would expect to see here that is missing.
Any assumption the argument depends on that has not been stated.
THE WORK:
[paste]

That prompt is useful, but understand its limit. It generates the list of things to verify. It does not verify them. The verifying is your job, using your knowledge and real sources.

When you have not got time to check everything, check a small amount properly rather than skimming all of it. A spot check that goes deep on three claims tells you far more about the reliability of a document than a full read-through that goes deep on none.

I only have twenty minutes to check this document, so I cannot verify all of it.
Pick the three claims where being wrong would cause the most damage, given that this is going to [say who reads it and what they will do with it]. Rank them by consequence, not by how likely they are to be wrong, and tell me the single fastest way to check each one.
THE DOCUMENT:
[paste]

Building a reputation for catching things

This part is quieter than people expect. The colleague everyone trusts is rarely the loudest critic in the room. They tend to do three things consistently.

They catch things early and privately. Flagging a problem to one person before a document circulates is a favour. Flagging it in a meeting after twenty people have read it is a performance. Same catch, very different reputation.

They say what they checked and what they did not. A note saying that you verified the figures and the regulatory references but did not check the customer numbers is enormously more useful than a general thumbs up, and it protects everyone including you.

They are specific about the risk. Not that something feels off, but that this figure comes from a source that was updated last year, and section three depends on it.

Help me write a short, low-drama note to a colleague flagging a problem I found in their draft. I want it to be specific about the risk, easy to act on, and not make them feel caught out.
Keep it under 120 words. Give me two versions: one for a peer and one for someone more senior than me.
WHAT I FOUND:
[describe the issue and why it matters]

The uncomfortable implication

If being the checker requires depth, then depth is the investment, and depth is slow. There is no version of this where you skim a tool tutorial and arrive. The good news is that slow investments are exactly the ones other people are not making, which is what makes them worth something.