The Second Look Makes the Expert

Case of the Week, Min Wu, PhD · ai-public-health.com

I bought a used car a few years ago. Low mileage, and it looked brand new — the kind of shine that makes you stop reading the rest of the listing. A short while after I drove it home, I started hearing a noise I couldn't place. I dug around and eventually found out the car had been in an accident before I ever saw it, and the internal structure never fully recovered — none of it visible from the outside. I sold it back at a loss and spent two years without a peaceful mind about the whole thing. If I'd actually known cars, I'd have popped the hood before I signed anything.

Key Takeaways

• Expert-In is expertise doing two jobs, not one: framing the question well, and coming back to check what the answer gave you.
• A polished, fluent answer — like a freshly detailed car — tells you nothing about whether it's actually sound underneath.
• Skipping the second half of Expert-In doesn't just risk a wrong answer. It risks accepting a wrong answer with full confidence, because nothing about it looked wrong.

I wrote last week about Fluent is not the same as right, and one line in it has been sitting with me since: expertise matters more with AI, not less. I've been turning that over, and I think it actually splits into two separate claims that get collapsed into one. This essay is about pulling them apart.

Two Jobs, Not One

We tend to think of expertise as something you bring to the front end of a task — the knowledge that lets you ask a sharper question, give better context, frame the problem well. That's real, and it matters. But it's only half the job.

The other half shows up after the answer arrives, not before. It's the expertise that lets you look at a fluent, confident response and actually test it — check it against what you know, notice the part that got quietly skipped, catch the assumption that doesn't hold. Framing the question well doesn't do that work for you. It's a separate skill, applied at a separate moment, and it's the one that's easiest to skip because by the time you're reading the answer, it already looks finished.

I've started calling this whole loop Expert-In: the expertise you put into the question, and the expertise you put into checking the answer, are the same currency spent twice. Most people only budget for the first spend.

One scope note before I go further: this is about the judgment-call questions — the ones without an answer key. If your task is closed enough to check against reality, you don't need Expert-In, you need a reality check. Expert-In is for the harder territory last week's post called "it depends."

The Shine Isn't the Structure

That used car was a masterclass in this, just not the kind I wanted. The finish was genuinely excellent. Nothing about the paint or the interior gave away what turned out to be underneath — and to be fair, that's not unusual; a clean exterior and sound structure aren't always the same purchase. The polish of the presentation and the soundness of the car were simply two different things, and I only had the tools to evaluate one of them.

AI answers can work the same way, and I don't think that's any one tool or company's fault — it's closer to a structural fact about fluent text. A well-organized, confident paragraph is a presentation layer. It tells you the system can produce something that reads like an expert wrote it. It doesn't tell you whether the internal structure — the actual reasoning, the facts it leaned on, the part of the question it may have quietly skipped — is sound. Those two things can move independently, and a novice has no built-in way to tell them apart. That's not a knock on novices, and it's not an accusation against the tool. It's just what happens when the only signal you're using is "does this sound right," and sounding right was never designed to certify being right.

What Popping the Hood Actually Looks Like

So what does the second job of expertise actually involve, in practice? A few things, none of them exotic:

Checking the answer against something outside the AI — a source you trust, a colleague, a fact you already know cold. Asking the AI to show its reasoning, not just its conclusion, and reading that reasoning skeptically rather than as confirmation. Testing an edge case the answer didn't mention, the way you'd listen for a noise a walk-around inspection wouldn't catch. None of this is glamorous. It's the equivalent of the cold-start idle and the OBD scan — unglamorous checks that only mean something if you know what a bad reading looks like.

That last part is the catch. The second half of Expert-In takes its own expertise, separate from whatever you needed to ask the question well in the first place. Which means the students I worry about most aren't the ones who don't know how to prompt. They're the ones who can prompt beautifully and have nothing to check the answer against once it arrives.

Expert-In means twice, not once.

Observer's Insight

Last week I ended without a clean answer for the student who's genuinely on their own with a judgment call and no expert in reach. I still don't have one. But I think this essay sharpens the question rather than answering it: it's not enough to say "go find someone who knows the subject." What that person actually offers isn't the right answer — it's the audit. The habit of popping the hood instead of admiring the paint.

I don't know yet how you teach that habit to someone who's never had a reason to distrust a good-looking answer. I suspect it isn't taught the first time something goes wrong — I suspect it's taught the second time, once the cost of skipping it is no longer hypothetical. That's a slower and more expensive way to learn it than I'd like for my students. I'd rather find the faster way, if there is one.

I'd be curious whether this two-jobs framing matches your own experience with AI tools — where has the fluency of an answer nearly gotten past your own check on it?

As a member of the Springer Nature Author Affiliate Program, I may earn a small commission from purchases made through this affiliated link to my book. Support the author by checking out my textbook, Artificial Intelligence in Public Health: https://tidd.ly/4mH9389