For most of my life, my writing read like a machine. Competent, structured, a little stiff — sentences assembled rather than spoken. I used to think that was a flaw. Recently I understood what it actually was: compensation. Language is, in a real sense, my second language; my first is pictures. So I learned to build sentences the way you'd build a system — logic carrying the load my verbal wiring couldn't. The result reads a lot like AI. And now I know why: AI and I solve problems the same way. Neither of us generates language natively. We both construct it. (My family finds it very funny that the machine writes like me, not the other way around).
I'm starting here because it's the whole reason I think the dominant story about AI and work is quietly measuring the wrong thing.
The leveling story — granted
You've seen the headline by now. Study after study finds the same shape: give people AI, and the lowest performers improve the most. Writers with weaker skills jump toward the level of the strong ones (Noy & Zhang, Science). Elite consultants see the gap between top and bottom narrow (the BCG "Jagged Frontier" study). The conclusion everyone draws: AI is a great leveler. It compresses the distribution. It lifts the bottom.
On the surface, that's exactly what the data shows. I won’t dispute it. I want to ask a deeper question about it.
The crack: Low performer — by what measure?
In every one of those studies, "performance" means output. Quality of the writing, prose, etc. The polish of the thing produced. And here's the problem hiding in plain sight: output capacity and thinking capacity are not the same thing. For a whole class of people, they come apart — hard.
Some people see the answer clearly and can't get it onto the page. The bottleneck was never the idea — it was the translation. The drafting. The turning-of-vision-into-words tax that some pay on every sentence. Measure those people on standard conceptions of output, and they score low — not because they think poorly, but because the instrument is reading the wrong channel.
Now watch what AI does. Noy and Zhang found something more specific than "AI helps": it mostly substitutes for effort, not skill, and it shifts the human's work toward idea-generation and away from rough-drafting. Read that twice. AI eats the drafting. It removes the exact tax that was suppressing the score. So of course, the "low performers" jump — for some of them, the low score was never the ceiling. It was the toll.
The evidence the field already had. This isn't a hunch. The talent research has known for fifty years that there's a whole dimension of ability our instruments systematically miss.
Lubinski, Benbow, and Wai at Vanderbilt have tracked thousands of people across decades. Their finding: spatial reasoning — the capacity to hold and turn a whole system in your head — is a distinct ability, separable from verbal and mathematical reasoning, and it predicts who builds real things (patents, technical breakthroughs) over and above what verbal and math scores capture. And yet it's almost never measured — not in school, not in admissions, not in hiring. A large pool of brilliant people goes unidentified. And the spatially strong who happen to be weaker verbally tend to underachieve in traditional settings, precisely because their strength is the one nobody scores.
Sit with what that means. The instruments we use to sort people — tests, essays, interviews, performance reviews — over-weight the verbal/sequential channel and under-weight the systems/spatial one. So, a specific kind of person gets mislabeled for years: high conception, real building instinct, throttled output. They don't climb the linear ladder. They route around it — into design, into operating, into founding.
The reframe: So here's the reframe, and I'll hold it as a hypothesis, because that's honestly what it is. What if AI isn't only leveling the field? What if, for a meaningful slice of people, it's un-maskingthem?
Not lifting low-capacity workers to adequacy — though it does some of that too. But releasing high-capacity people whose output was throttled by a translation tax the old instruments scored incorrectly. If that's true, the leveling statistic has an un-masking effect folded inside it that the aggregate numbers can't see — because no study has yet measured conception separately from articulation and asked which people the AI boost lands on.
There's a version of this everyone already accepts. AI is a multiplier: feed it bad data, broken processes, sloppy thinking, and it doesn't fix the mess — it scales it. Garbage in, garbage magnified. Everyone nods at that. But a multiplier doesn't only work on garbage. Invert it. Point that same force at someone with genuine conception — real systems thinking, the ability to see the whole board — who was only ever capped by the tax of getting it into words, and you don't get magnified mess. You get magnified signal. The same force that scales dysfunction scales latent talent. We've only been talking about one direction of it.
That study doesn't exist, yet. I want to be precise: this is not established. It's a falsifiable claim waiting for the right experiment — measure the gap between what a person can conceive and what they can articulate, then see whether AI's boost is largest exactly where that gap is widest. Until someone runs it, "we haven't measured it" is not the same as "it isn't real." And it isn't the same as "it's true," either. It's open. I think it's the most interesting open question in the whole conversation.
What to do before the study exists. You don't have to wait for the evidence to act on it, because the practical move is the same either way: Stop treating output polish as if it were capacity. They were never the same, and AI just made the difference enormous — because the moment a tool can carry the articulation tax, the polished-output advantage you were screening for evaporates, and what's left is the thing you weren't measuring: conception, judgment, the ability to see the whole board. (Which, conveniently, is also the thing AI can't do — and the thing that separates the companies getting real ROI from the 95% that aren't. But that's the next piece).
If you run a team, this is a hiring and a measurement problem before it's a technology problem. The person whose memos were always a little rough but whose instincts were always a little right? Re-evaluate them now. The tax that was hiding them just dropped to zero.
The window: I opened with my own wiring because it's the clearest example I have — but the point was never personal, and it isn't about one personality type. It's about what your company optimizes for.
The strategic version is simple. AI devalued every screen that measured polish and left untouched every problem that needs judgment (The most valuable skill in the AI era). Most companies will spend the next two years admiring the productivity bump and miss the move underneath it: the people their instruments quietly filed as average are now the most mispriced talent on the board. That's not an HR footnote — it's an arbitrage window, and it closes the moment everyone else notices.
First movers here don't just adopt AI faster. They inherit a workforce the rest of the market is still scoring wrong.
Sources
- Noy, S. & Zhang, W. "Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence." Science (2023). AI raised average productivity and compressed the distribution, helping lower-skill writers most; it substituted largely for effort and shifted work toward idea-generation and away from drafting.
- Dell'Acqua, F., et al. "Navigating the Jagged Technological Frontier." Harvard Business School / BCG working paper (2023); forthcoming in Organization Science. Inside AI's frontier, large productivity and quality gains, largest for lower performers; the study also documented compressed idea diversity across consultants.
- Wai, J., Lubinski, D., & Benbow, C. P. "Spatial Ability for STEM Domains." Journal of Educational Psychology (2009); Kell, H. et al. "Creativity and Technical Innovation." Psychological Science (2013). Spatial reasoning is a distinct, systematically under-measured dimension of ability, with incremental validity for predicting patents and publications beyond verbal and mathematical measures.
I work with scale-ups and small companies in exactly this transition — bringing senior, AI-native operating judgment in fractionally, so you get the whole-board view without the full-time cost. If your market is moving faster than your operating model, let's talk.