When explanation becomes free and any fact is one prompt away, the value of learning does not fall to zero. It moves. It relocates from acquiring information — now abundant, instant, and nearly costless — toward the things a model cannot hand you across the chat window: judgment, taste, tacit knowledge, and enough real expertise to tell a correct answer from a confident wrong one. The question every learner is quietly asking — why learn anything if AI already knows it? — has a clean answer. You are confusing access to explanations with having a model, and you are forgetting that someone in the room still has to know whether the machine is right.
The seductive conclusion fails in two distinct places, and both matter.
Access to an explanation is not the having of a model
The first failure is a category error about what learning is. An explanation is a sequence of sentences. Understanding is a structure in your head — a compressed, navigable model that lets you run several inferential steps without stopping to look anything up. You cannot outsource the building of that structure to a tool that hands you finished sentences, any more than you can get fit by watching someone else lift. The reps are the point.
Learning science has been blunt about this for decades. Robert Bjork's work on "desirable difficulties" shows that the conditions which make learning feel easy and fast — rereading, being handed the answer, massed practice — produce fragile, short-lived retention, while the conditions that feel effortful and slow — retrieval practice, spacing, having to generate the answer before you see it — produce durable understanding. The struggle is not friction on the way to learning. The struggle is the learning. A tool that removes it by explaining everything on demand is optimizing the exact variable Bjork's research says you should leave alone.
This is the trap hiding inside a genuinely wonderful capability. Benjamin Bloom's 1984 "2 sigma" result found that students given one-on-one tutoring outperformed conventional classroom students by around two standard deviations — a staggering effect, and the reason a tutor for every person is not hype. But Bloom's tutors worked by inducing productive struggle: diagnosing the gap, then prompting the student to close it themselves through mastery learning and corrective feedback. A tutor that simply answers is not the thing that produced the two sigma. The delivery mechanism is now free; the pedagogy is not automatic, and the default mode of "explain this to me" quietly skips it.
So free explanation, used passively, can leave you knowing of a great many things and understanding almost none of them. That is a real regression, not a wash.
You need internalized expertise to judge the machine
The second failure is more consequential, because it holds even for people who use AI well. To extract value from a system that produces fluent, authoritative-sounding output, you need enough internalized expertise to evaluate that output. Without it, you cannot distinguish a correct answer from a confident wrong one — and current models produce both in the same even, assured register.
This is the deepest reason learning still matters, and it inverts the naive picture. The naive picture says: since the AI knows the domain, you don't have to. The truth is the opposite. Precisely because the AI will hand you a plausible answer every single time, including when it is wrong, you need the domain knowledge to catch it. Novice and expert receive the identical confident paragraph. The novice is fooled; the expert catches the error. The only difference between them is the internalized model the naive picture told you not to bother building.
The asymmetry that makes this workable is that verification is cheaper than origination. You do not need to derive the answer from scratch to check it — you need enough of a model to run sanity checks, recognize the failure signatures, and know which claims deserve a second look. A seasoned engineer who couldn't reproduce a particular algorithm from memory can still see in one glance that a proposed solution is quadratic where it should be log-linear. A physician reading a differential can feel the one diagnosis that doesn't fit the labs. That felt wrongness is compressed expertise — thousands of prior cases collapsed into a fast check. It is the most valuable thing you own in an AI workflow, and it is only available to people who did the slow work of building it. I wrote about the endpoint of this trend in When Competence Is Free, Taste Is the Last Moat: when the model can generate a thousand competent options, the scarce and defensible skill is the judgment that knows which one is right. Verification-grade expertise is the near cousin of taste — the knowledge that lets you grade the machine instead of trusting it.
There is a useful historical analogy, and it's worth getting right. After Deep Blue beat Kasparov in 1997, Kasparov championed "advanced chess," later generalized as centaur chess — human and engine playing as a team. For a period, human-plus-machine pairs beat both unaided humans and, in some formats, unaided engines. The lesson people take is "collaborate with the AI." The sharper lesson is what the human contributed: not raw calculation, which the engine did better, but judgment about which lines to explore, when to trust the evaluation, and when to override it. The centaur's edge came from a human who understood chess deeply enough to direct and second-guess a stronger calculator. A weak player centaured with the same engine added nothing. Access to the engine was necessary and nowhere near sufficient. The honest caveat: as engines grew overwhelmingly strong, the human's marginal contribution shrank — a warning I'll return to.
What becomes more worth learning, and what becomes less
If information acquisition is depreciating and evaluation is appreciating, the rational move is to rebalance where your finite learning hours go. I made the general case for treating skills as a portfolio with different half-lives in Your Skills Have a Depreciation Schedule. Are You Rebalancing?. Here is the specific reallocation that "AI can explain anything" forces.
| Depreciating (offload or thin out) | Appreciating (invest deliberately) |
|---|---|
| Rote facts, easily retrieved | Judgment: choosing what's worth doing, good vs. great |
| Memorized syntax, API signatures, reference figures | Verification-grade domain depth: enough to direct and catch the AI |
| Procedural knowledge with a stable, lookup-able answer | Tacit, experiential knowledge not in any corpus |
| Information about a field | Meta-skills: how to learn, specify, and ask |
Four things become more worth learning.
Judgment and taste. Deciding what is worth doing at all, and telling the merely good from the great, is a selection function the model cannot supply because the training signal barely exists. The web is full of competent artifacts and nearly empty of trustworthy verdicts about which was the right call in which context. Choosing the problem, setting the bar, and knowing when to stop are yours.
Deep domain expertise, specifically to the depth of verification. Not encyclopedic recall — the model has that — but enough structural understanding to aim the tool at the right question and catch it when the answer is wrong. This is the difference between consuming AI and commanding it.
Tacit knowledge — the kind that isn't in any corpus. Michael Polanyi's line, "we know more than we can tell," names the category: the surgeon's hands, the negotiator's read of a room, the operator's feel for when a system is about to break. It is learned by doing under real stakes and feedback, not by reading — which makes it structurally unavailable to a model trained on text, and equally unavailable to you if you only ever read the machine's explanations instead of doing the thing.
Meta-skills that compound as tools change. How to learn a new domain fast, how to specify a problem precisely enough that someone — or something — else can execute it, how to ask the question that actually resolves your uncertainty. Prompting is a temporary, shallow instance of the durable skill: decomposing a fuzzy goal into a crisp specification. That skill was valuable before AI and grows more valuable as the executor on the other end gets more capable.
And one thing becomes genuinely less worth memorizing: cheaply retrievable, stable facts where a wrong recall is easy to catch. Exact function signatures, boilerplate, tax tables, the reference material you'd look up anyway. Offload it.
The critical caveat — the one that keeps this from collapsing into "don't learn facts" — is that you cannot reason with a database you have to query on every step. Fluent thinking in a domain requires a resident scaffold of load-bearing knowledge: the core structure you reason with, held in working memory, not the reference material you reason against. If every third inferential step forces a lookup, you never build the chains of reasoning that produce insight, and — worse — you lack the internal model to know whether the retrieved fact is even relevant. Memorize the skeleton. Offload the footnotes. The skill is knowing which is which.
The strongest counterargument, stated fairly
The best version of the opposing case is not "AI knows everything, so learning is dead." It's the centaur warning made general: if the machine keeps getting better, the human's marginal contribution shrinks toward zero, verification included. In chess, the centaur era ended — top engines became so strong that a human in the loop added noise, not signal. If domain models cross the threshold where they are more reliable evaluators than any human expert, then even verification-grade expertise depreciates, and the whole "learn to judge the machine" edifice goes with it.
I take this seriously, and I'll label my position as a bet rather than a fact. Two reasons I think the verification premium is durable for a long time even so. First, chess is a closed system with a perfect, cheap reward signal; most valuable human domains are open, contextual, and have sparse, expensive feedback — exactly the regime where automation crawls rather than sprints. Second, even a superhuman evaluator has to be pointed at the right question, and its output has to be owned by someone accountable for the consequences — and specifying the right problem and bearing responsibility for the answer are not tasks a more accurate oracle removes. Whether that holds through a genuinely sharp capability jump, I don't know, and anyone who claims certainty about takeoff speed — fast or slow — is bluffing past the evidence. That's the uncertainty, stated plainly.
Here is the concrete move, whatever the timeline. Stop spending your scarce learning hours acquiring information the model already serves for free. Spend them on what it can't hand you: reintroduce the difficulty by predicting before you ask, do enough by hand to earn the intuition, and get your judgment graded against someone better. The learner who wins the AI era is not the one who memorized the most. It's the one who can look at a confident answer and know.