A marketing team used to spend two weeks producing one campaign concept. Now they can generate a dozen before lunch. A designer who labored over three logo options can conjure a hundred. A writer stares not at a blank page but at six competent drafts.
Notice what just happened to the scarce thing. It used to be the production—the hours, the craft, the effort of making the thing at all. That's now nearly free. What's scarce is the answer to a harder question: which of these is actually any good—and are you willing to put your name on it?
Most people are asking what AI makes cheap. The more useful question is what it makes valuable.
When something gets cheap, its complements get precious
In Prediction Machines, the economists Ajay Agrawal, Joshua Gans, and Avi Goldfarb make a deceptively simple point: AI is, at bottom, a drop in the cost of prediction. And when the price of one thing collapses, whatever it can't do for itself becomes the bottleneck—and the prize.1
Think of photography. When phones made taking a picture essentially free, we didn't get more great images so much as vastly more images—and the scarce skill became the eye to know which single frame was worth keeping. Cheap capture; expensive judgment. Generative AI does the same to knowledge work: it has collapsed the cost of producing plausible content—text, code, analysis, design—so the value shifts to the thing it can't supply, the judgment to know which output is actually good. Agrawal and his colleagues put it precisely: judgment is what you need exactly when the goal can't be fully written down.1 That describes most of the work that matters.
Why the machine can't judge its own work
Software was the first domain AI transformed because code has an answer key: a unit test passes or it doesn't, so the machine can generate, check, and iterate its way to a good answer (I've written about this as the measurable-output test).
Most work has no such key. Cassie Kozyrkov calls it the world of "many right answers": ask AI for a marketing plan and you don't get the answer, you get a distribution of plausible ones.2 There is no test the machine can run to pick the best, because "best" depends on context, strategy, and taste that were never written down. The AI can hand you the dozen. It cannot tell you which one to run—and it certainly cannot decide to stand behind one. Someone has to.
Skip that step—ship whatever the machine produced without the judgment—and you get workslop: output that looks finished, reads fluently, and turns out to be hollow the moment an expert examines it. Workslop is simply what production looks like with the judgment left out.
Judgment becomes the job
The job that's left is narrow and hard: out of many plausible options, choose the one worth pursuing, and own that choice. You can generate ten marketing plans; you're going to commit to one.
But choosing is only the start, and it's the part the "pick the best of ten" story misses. An expert rarely runs even the best option as-is. They treat it as a starting point—the AI got them 80 or 90 percent of the way there, and often surfaced a direction they wouldn't have thought of—and then they edit and refine it with the expertise the model doesn't have. The generating is shared; the shaping and the final call are theirs. That last stretch, closing the distance between "plausible" and "right," is where the judgment actually lives.
That work is moving to the center of most jobs. The analyst's value shifts from producing the deck to deciding which finding deserves the board's attention and sharpening how it's framed. The lawyer's, from drafting the memo to judging which argument will persuade this particular judge—and rewriting it until it does. Making the options got cheap. Choosing the right one, refining it, and committing to it is the job now.
Trust becomes the moat
Scale that up and it turns into competitive advantage. When anyone can generate a professional-looking anything in seconds, platforms and markets fill with output that looks finished and no one has vetted—and the scarce, valuable thing becomes a source you trust to have already done the judging.
Watch it happen in real time. Apple's App Store took in nearly 560,000 new apps in the first half of 2026—approaching the total for all of the previous year—as "vibe coding" let non-developers ship apps in days. The flood is straining Apple's human review, and scam apps are slipping through.3 Steam has the mirror-image problem: thousands of AI-generated "asset flips" now bury the genuinely good games, and Valve has resorted to algorithmically splitting the store into the part people see and, in effect, the landfill.4 In both cases the bottleneck isn't making apps or games—that's trivial now—it's the trusted judgment that separates what's worth your time from what isn't. That judgment doesn't scale easily, which is exactly why it's becoming valuable.
What it means for you
This reframes what "getting good at AI" even means. The temptation is to measure it in output—how much faster you can produce. But production is the part that just got cheap. The skill worth building is the other one: the judgment to look at what the machine made and know, quickly and correctly, what's excellent, what's wrong, and what to do next.
It's also why I think AI widens the gap between strong and weak performers rather than closing it. Give a novice a dozen AI drafts and they can't tell which is best, so they ship the plausible one—and quietly produce slop. Give an expert the same dozen and they keep the good, cut the rest, and rework it into something none of the drafts was on its own. Same tool, opposite result—because the expert brought the judgment the tool can't supply.
The scarcity moved
AI made producing things almost free. It did not make choosing well free—if anything it made choosing harder, by handing us more plausible options than we can sort. So the value moved: from the hands that make to the judgment that picks, from volume to discernment, from the output itself to the person willing to stand behind it.
The people and firms who win the next few years won't be the ones who generate the most. They'll be the ones who can look at endless plausible output and say, reliably, this one—and here's why.
Footnotes
-
Ajay Agrawal, Joshua Gans, and Avi Goldfarb, Prediction Machines: The Simple Economics of Artificial Intelligence (Harvard Business Review Press, 2018); see also their "Prediction versus judgment" work: https://www.nber.org/system/files/working_papers/w24626/w24626.pdf ↩ ↩2
-
Cassie Kozyrkov on AI producing distributions of valid outputs that require human judgment to select among ("many right answers"): https://decision.substack.com ↩
-
On the App Store flood and review strain: Josipa Majic, "The Apple App Store Is Flooded With AI Slop And Legitimate Developers Are Paying For It," Forbes (March 2026): https://www.forbes.com/sites/josipamajic/2026/03/24/the-apple-app-store-is-flooded-with-ai-slop-and-legitimate-developers-are-paying-for-it/ ; and 9to5Mac on the ~84% surge in new-app submissions: https://9to5mac.com/2026/04/06/app-store-sees-84-surge-in-new-apps-as-ai-coding-tools-take-off/ ↩
-
On AI "asset flips" and discoverability on Steam: PC Gamer, https://www.pcgamer.com/steam-sends-90-low-effort-asset-flips-and-bootleg-games-off-to-the-great-trashcan-fire-in-the-sky/ ; and Aftermath, https://aftermath.site/steam-next-fest-demos-ai-disclosure-algorithm/ ↩