Watch a junior analyst ask AI for a market-entry plan for the first time. They type a few sentences—we're a mid-size coffee brand, should we launch in Southeast Asia?—and a full strategy unspools onto the screen: market sizing, a competitor breakdown, a phased rollout, a budget, a confident revenue projection at the bottom. Ninety seconds. Their jaw hits the floor. This is amazing.
And it is. Given what that analyst has seen so far, the reaction is exactly right.
Here's what I want to admit up front, because it complicates the easy version of this story: it happens to me too. I've spent years building business cases and picking apart other people's, and the first time I watched AI produce a whole strategy in one shot, my jaw dropped just like theirs. Even now, in an area I know well, a good AI output can stop me cold before I've thought about it. The wonder is real, and it's close to universal. So the difference between me and the analyst isn't that I'm immune to being impressed. It's that a beat later, I look again—and I have the equipment to see what the second look turns up.
That second look is the whole subject of this essay.
Two questions hiding in one word
When the analyst says "amazing," the word is quietly doing two different jobs.
One is amazing compared to me—compared to what I could have produced myself, in the time I had, with the skill I've got. On that scale the plan is genuinely astonishing. Someone early in their career could not build that in ninety seconds, or in ninety minutes.
The other is good compared to a seasoned pro—compared to what someone at the top of their field would actually put their name on and hand to a client. That's a different scale entirely, and on it the same plan might be mediocre or quietly broken.
The trouble is that for the novice these two readings feel like the same reading. They only own the first scale, so "amazing versus me" and "good versus an expert" collapse into a single verdict: amazing, full stop. The mistake isn't being impressed. The impression is accurate. The mistake is not noticing that you answered an easier question than the one that matters.
Why you can only see what you've learned to see
Here's the uncomfortable mechanism underneath. You judge "amazing" against everything you've seen before—your reference set. A novice's reference set is small, so almost anything competent clears it. An expert's reference set is enormous, so the same output gets measured against the best work they've ever encountered, and the flaws light up.
Take that AI market-entry plan. To a beginner it looks complete because it has all the parts—a market size, a competitor grid, a rollout timeline, a confident revenue line at the end. What a trained eye sees is different: a plan whose entire case rests on one unexamined assumption—that you'll win, say, 10% of the market in year one—with nothing underneath it; adoption and growth numbers nobody justified; a five-year projection carried to the dollar on top of a guess. It names every section a real plan would have and skips the reasoning that makes a plan worth betting on—no serious read of why the incumbents would let you win. The beginner can't see the missing layer—not because they're careless, but because seeing it is the expertise. You cannot spot the absence of something you've never learned to expect.
This is why AI amazes you in inverse proportion to how much you know. The poet is the one unmoved by the AI poem; the rest of us are dazzled. When OpenAI's Sam Altman shared a ChatGPT-written short story he found sublime, the novelist Dave Eggers read the same text and laughed—it was, to him, obviously hollow.1 Same words, opposite verdicts. Altman isn't a fool; he's just not a writer, so how would he know?
Which points to the real principle, and I want to state it carefully because it's easy to make it sound like snobbery. It isn't that you lack the right to judge. It's that judging quality is an earned capacity, not a granted one. AI hands the production to everyone for free. It cannot hand you the eye. That still has to be earned the old way—by getting good enough at the thing that you can finally see what "good" means. Production got democratized. Evaluation did not.
The bluffer believes his own bluff
Now watch what happens when the person who can't tell the difference is you, evaluating yourself.
Bloomberg recently told the story of a grant writer who used AI to get through his interviews, landed the job, turned out to be unable to do it, and was gone within a couple of months.2 It's tempting to file him under cynics—a con man who knew he was unqualified and gamed the test. I think the truth is sadder and far more common: he believed it. AI helped him sound expert in the interview, and he read his own fluent answers the way the analyst reads the plan—this is amazing, I can do this. The evaluation gap doesn't only mislead you about the machine's work. It misleads you about your own. He never learned to see the distance between sounding capable and being capable, so he mistook the first for the second.
That's the illusion of expertise, and it's a cousin of workslop: workslop is output that looks like work and isn't, and this is a worker who looks capable and isn't—most of all to himself. The essay's warning runs in two directions from here. To the junior person tempted to be the bluffer: AI can get you the job, but the job eventually asks you to actually do the thing, and AI cannot hand you the competence you skipped building. To the manager on the other side of the table: a sincere self-deceiver with good AI will walk straight through a hiring process built to catch a fraud, because he isn't performing confidence—he genuinely has it. It’s just that the confidence is built on a foundation of sand.
When average is exactly enough
None of this means AI's average-grade output is worthless. Often it's a gift.
Picture a solo founder who can't afford to hire anyone. AI takes them from knowing nothing to roughly average in a domain outside their own—a serviceable contract, a decent logo, marketing copy that reads fine. Against the realistic alternative, which is nothing, or a worse do-it-yourself attempt, average is a real win. The output doesn't have to be expert. It has to clear the bar for the job, and here the bar is low and the cost of being a little wrong is small.
That's not a counterexample to everything above—it's the discipline in action. The founder who thinks "I need serviceable copy and serviceable is fine" has read the stakes correctly. That reading is the skill. The danger was never average work; it's average work mistaken for expert work in the places where the difference is expensive and no one in the room can see it. The useful question is the same one I keep coming back to: not can AI do this? but how good does this actually need to be, and what does it cost me when it's a little off?3
Keep the wonder. Add the second look.
So let the AI drop your jaw. Mine still does. The wonder isn't the problem—it's honest, and the tools have earned it.
The problem is stopping there. If you're a novice in a field, treat your own amazement as information about you, not just about the output: the more easily something impresses you, the less you may know about what excellent looks like. The fix isn't to become a cynic. It's to go build the reference set—the trained eye that lets you look twice and know what the second look means. That's the one thing AI still can't do for you, which is exactly why it's the thing worth having.4
AI is amazing. Whether it's any good is a different question—and learning to tell the two apart is most of an education.
Footnotes
-
Dave Eggers, interviewed in the Financial Times, on his reaction to the AI-written short story Sam Altman praised as sublime: https://www.ft.com/content/1c833e7b-3cba-4b87-8b45-33d6c2bd468e ↩
-
Bloomberg Businessweek, "AI Tools Can Help Job Hunters Cheat on Interviews and Coding Tests" (July 2026): https://www.bloomberg.com/news/articles/2026-07-14/ai-tools-can-help-job-hunters-cheat-on-interviews-and-coding-tests ↩
-
The deployment question I develop in Capable, but Not Reliable. ↩
-
The value-side of this argument—why the trained eye becomes scarce and therefore valuable—is The Judgment Premium. ↩