You can't ask what you don't know to ask

Here's the quiet catch in "AI democratizes expertise": to get expert output, you have to ask an expert question. And if you already knew the expert question, you wouldn't need the expertise.

Watch a specialist and a beginner use the same model on the same task. The specialist gets something great — because they know what to specify, what constraints matter, which edge cases to name, what "good" even looks like. The beginner types the obvious question, gets a fluent, generic answer, and has no idea what they didn't ask for. Same tool. Opposite outcomes. The gap isn't the model. It's what each person knew to put into it.

This isn't a vague observation. It's a named problem, several of them stacked, and once you see the names you can't unsee the trap.

The names for the trap

You can't perceive the gap you're missing. The philosopher's term is meta-ignorance: not knowing what you don't know. You've heard its pop version, "unknown unknowns." You cannot form a question about a thing you don't know exists, so the most important context is exactly the context you'll never think to provide.

The people who need to ask well are the least able to. That's the Dunning-Kruger effect, stated properly: the same missing skill that makes you perform poorly also removes your ability to notice you're performing poorly. Dunning called it being "doubly cursed." Applied to AI: the person who most needs a good question is the least equipped to know theirs is bad.

Wanting the answer is what stops you specifying it. Information science has the cleanest version. Nicholas Belkin's Anomalous State of Knowledge: you go looking for information because there's a hole in what you know — but that same hole means you can't say precisely what you're looking for. If you could state it exactly, you'd half-know it already. Query quality is capped by the very knowledge the query is meant to acquire.

Put those together and you get the mechanism behind the last thing we wrote about — how AI silently guesses who you are when you leave a gap. The blog before named the model's half of the failure. This is the human half: you don't leave gaps on purpose. You leave them because you can't see them.

Why this makes AI widen gaps instead of closing them

The optimistic story is that AI lifts everyone. The structure says otherwise.

Educational psychology has a finding called the expertise reversal effect: support that helps a novice can actively hold back an expert, and vice versa. The expert brings the missing context and the sharp question; the tool amplifies them. The novice can't, so the tool fills the blanks itself — quietly, from how the question sounded — and hands back something plausible and average.

So the same tool multiplies what you already have. That's a Matthew effect — the rich get richer. The more you know, the more you can extract; the less you know, the less you can even ask for. AI as a competence multiplier, not a competence substitute.

The finance study we've been citing is this on a graph. Researchers had a thousand real people ask an AI for money advice. The gap between good and bad lifetime outcomes came mostly from the demand side — what people knew to include — not from the model. Only 6% thought to mention an emergency fund. Those omissions compounded into real differences in retirement wealth. Nobody withheld the context. They didn't know it was the context that mattered.

The fix is not a smarter model

This is the part worth sitting with, because it's counterintuitive. A better model does not close this gap. A better model is more capable of running with your unstated assumptions, more fluent at making a thin question sound answered. Raw capability, if anything, widens the articulation gap, because it raises the ceiling for people who can ask and leaves the floor exactly where it was for people who can't.

If the problem were "answers are wrong," you'd wait for GPT-6. But the problem is "people can't ask what they don't know to ask." That's not a model problem. It's a scaffolding problem — and scaffolding is a thing you can build today.

Borrowed articulation

Here's the reframe. What a novice is actually missing isn't intelligence. The intelligence is right there in the model, cheap and waiting. What they're missing is the question an expert would have asked. The set of things to specify. The constraints that matter. The context you don't know is context.

So encode that. A good prompt template is not a shortcut for typing — it's a captured expert question. It carries the fields a specialist would have insisted on, the assumptions they'd have made explicit, the checks they'd have run. When a beginner fills it in, they're not becoming an expert. They're borrowing an expert's articulation — running the question they couldn't have written themselves.

That's the whole idea. A prompt library, a structured skill, a saved context profile — underneath, they're all the same move: someone who knew what to ask, packaged so someone who didn't can ask it anyway. It turns meta-ignorance from a fatal flaw into a solved onboarding step, because the missing questions come pre-loaded.

This is what Korvai is really for. Not "save your prompts" — that's the mechanism. The point is to make the expert's question portable: to close the articulation gap with structure instead of waiting for people to somehow become experts before they're allowed good output. The model already democratized the answers. What's left is democratizing the questions.

You can't ask what you don't know to ask. But you can run a question someone else knew to write. That's the difference between AI making the gap wider and AI finally making it smaller.