What AI can't learn from you

For most of the last century, expertise was an appreciating asset. You trained in something, earned the credential that certified it, and spent decades letting experience compound on top. The mechanical engineer of 1985 was, in the main, still the mechanical engineer of 1995. The entire architecture of careers rested on one quiet assumption: that a body of hard-won, learnable expertise holds its value long enough to be worth acquiring.

That assumption is the thing AI actually breaks—and I don't think we've been honest about what that means, because the honest version is inconvenient.

The usual framing is that skills now have a shorter half-life, that things simply decay faster than they used to. True, but it undersells what's happening. Every previous wave of automation ate manual work and routine work—the tasks we were mostly happy to hand off. This wave is different in kind. It goes after learnable 'expert' skill: the analysis, the drafting, the diagnosis, the modelling, the structured reasoning.

Which is to say, precisely the thing a credential exists to certify. When you earn a qualification, you are certifying that you have absorbed a transferable body of expertise. That category is now the most automatable one we have—because a model can absorb the same body instantly, at almost no marginal cost, and never forget a word of it.

So, this isn't a story about skills decaying a bit faster. It's a story about the machine learning the expert's job at the same time the expert does and then doing it for free. Reskilling, seen in that light, can start to feel less like getting ahead of the curve and more like sprinting on a treadmill that keeps speeding up.

The reflexive answer to all this is *more learning*. Learn continuously. Stack credentials. Reskill on repeat. It is the message the whole system—employers, institutions, people like me—has every reason to repeat, which is exactly why it's worth stating plainly that the simple version of it doesn't hold.

Because "just learn faster" quietly does something unfair. It takes a problem created by the pace of technology and the priorities of employers, and hands the entire cost of solving it—the time, the money, the anxiety — to the individual. It implies that if you fall behind, you didn't hustle hard enough, when the ground was moving the whole time. And it papers over an awkward fact: a great deal of learning doesn't stick.

Completion rates for self-directed online learning are notoriously low. Transfer from the classroom to the actual job is worse. More learning, on its own, was never going to be the answer to a problem this shaped.

None of which means learning stops mattering. It means we've been optimising the wrong layer of it.

If a machine can now hold the expertise, the value of a human acquiring it shifts. It moves off the expertise itself and onto everything around it: knowing which problems are worth solving in the first place; carrying the judgement to tell a good answer from a merely plausible one; holding the trust and context that let expertise actually get deployed inside a messy organisation; being able to work alongside systems that now know as much as you do about the narrow thing. The machine can hold the knowledge. It cannot yet hold the responsibility for what to do with it.

That is a genuinely harder thing to build than a syllabus of facts and procedures. Transferring perishable expertise—the exact layer AI is busy commoditising—is the thing our institutions do well. Building the durable human layer around it—judgement, discernment, the capacity to hold ambiguity and still decide — is the thing almost no one does well yet. That is the gap worth closing.

It is also why I don't buy the fashionable claim that degrees and deep credentials are finished. You cannot exercise judgement about a field you don't understand; taste is built on foundations, not vibes. The deep grounding a serious programme gives you is exactly what lets you supervise a machine that is fluent but not wise. What changes is the job that grounding does—not a gate you pass through once and then coast, but the floor you stand on while you keep rebuilding everything above it.

For professionals, the practical takeaway is less comforting than "learn how to learn," and more useful. Assume the specific expertise you're proud of is on a clock. Invest in the layer above it—judgement, trust, the ability to decide well when the answer isn't clean—because that's the part the machine isn't taking soon. And be wary of anyone who offers endless courses as the whole answer.

Knowledge, for the first time in history, is close to free. What remains scarce is the human capacity to decide what to do with it — which problem is worth the machine's fluency, which answer to trust, which risk to take. That capacity has always mattered. It's just that until now, it came bundled with expertise we had to earn the hard way. The bundle is coming apart. The people, and the institutions, who understand which half was ever the point will do more than adapt. They'll be building the one advantage that doesn't expire.

(Rohit Sharma is President, Consumer Business at upGrad)