Compressed or Destroyed

The same mechanism that makes AI a great teacher is the one that makes it a great way to stop learning. It depends entirely on what the learner does with the time the tool gives back, and the two outcomes look identical for about a year before they stop looking identical at all.

Start with the good version, because the evidence for it is strong. In the big study of customer support agents, access to an AI assistant raised productivity about fourteen percent on average, and about thirty-four percent for the newest and least experienced workers, with almost no gain for the veterans. The likely mechanism, the authors suggest, was that the tool captured the habits of the best agents and handed them to the beginners, moving new people down the experience curve faster than the job ever could on its own. That is apprenticeship compressed. The thing a good mentor does, transferring hard-won judgment to someone who has not earned it yet, done at scale and on demand. It is one of the most genuinely hopeful findings in all of this.

Now the other version of the same mechanism. If a beginner uses the tool to produce good work without ever doing the messy, unglamorous part themselves, they get the output of expertise without the formation of it. And the formation was never in the output. It was in the mess.

What the mess was for

Everyone who got good at something difficult got good through repetitions that felt like a waste of time while they were happening. The hours of debugging that slowly built an instinct for where bugs live. The bad first drafts that taught taste by being bad in instructive ways. The architecture you got wrong and had to live with, which is the only thing that ever teaches you to smell a wrong one early. None of that was efficient. All of it was the training, and it was disguised as drudgery the entire time.

A 2025 paper makes the structural case for taking this seriously, arguing that AI deskilling is not a personal failing but a property of the environment. When a system reliably does the work a person used to do, it can create what the authors call a capacity-hostile environment, one that erodes the very abilities that work used to develop, not because anyone chose to deskill but because the repetitions simply stopped happening. The danger is not that people get lazy. It is that the gym closes and nobody notices until they need the muscle.

Which means the hard managerial fact buried in all of this is that not all friction is waste. Some friction is training. This is the antifragile idea pointed at a person: Taleb's case that some things gain from disorder, that stress and difficulty are inputs a system needs and not just damage to remove. I have tried to build my own life and systems to work that way, and the same logic holds for a skill. Take away every stressor and the worker does not get stronger. They get more fragile. A company that strips out every inefficient repetition in the name of speed can save an enormous amount of time while hollowing out the pipeline that was supposed to produce its next generation of senior people. The repetitions a junior should keep are exactly the ones a spreadsheet would flag for elimination first, because from the outside developmental friction and pure waste look the same. Telling them apart is real work, and almost nobody is doing it on purpose.

The floor is rising

You can already see the shape this pushes careers into. When AI absorbs the bottom rungs of a job, the typing and the boilerplate and the basic lookup, the rungs do not get replaced with easier ones. They get replaced with harder ones. PwC's barometer is picking this up: AI-exposed entry-level jobs are increasingly asking for the senior, human-intensive skills, judgment and leadership and synthesis, earlier than people used to have to supply them. Other work on job postings finds the same shift, with demand rising for higher-order cognitive and social skills rather than the mechanical ones the tools now cover.

So the value of typing, boilerplate, generic summarizing, and mechanical translation between formats is falling, and the value of framing, taste, review, synthesis, domain judgment, and teaching is rising. The human either moves up into that work or gets hollowed out underneath it. The cruelty is in the timing. We are asking people to start their careers at an altitude that used to take a decade to reach, while removing the gentle lower slopes that used to get them there. The floor of the profession is rising, which is good for the work and brutal for whoever was standing on the old floor expecting to climb the normal way.

What I actually do about it

I should be honest about the limits of my own view here. I do run a company, but it is a tiny LLC, not the kind of place with a talent pipeline to measure or a bench of juniors to deskill, so the organizational version of this is something I have read about more than lived at that scale. At my size the picture nearly inverts. I do not have to worry about scaling. I pay for exactly one person's cognitive amplification, my own, I know precisely what it costs and how it gets used because I am the one using it, and I point it at every part of the business. For my own business that makes it very cheap labor and a very low cognitive cost, a no-brainer, and in most cases a write-off besides.

I happen to see the other end of it too. Inside a large employer I can only imagine the same bill looks enormous and the return genuinely questionable, which is the paradox at the heart of the whole problem. Same technology, opposite verdict. The only things that changed are the scale and the hidden ledgers that come with it: the coordination tax, the talent pipeline, the meetings that never get shorter. None of those reach a business of one, which is probably why I can see them so clearly from here. I am not drowning in them.

What I can speak to firsthand, then, is my own skills, where the deskilling risk is the same mechanism, just smaller.

The accumulated feel I rely on, the thing that lets me sense where an agent is reliable and where I have to take the wheel, is nothing but the residue of years of doing the work the hard way before the tools existed, the kind of thing that compounds across a career and cannot be bought in a hurry. I did not download that. I cannot download more of it. And I have noticed that if I let the agent do a certain kind of thinking for long enough, the part of me that used to do it gets quieter, the way any unused thing does. So I keep some of it deliberately. I still do certain things by hand, not because the tool cannot, but because the doing is keeping a capacity alive that I am not willing to lose. The move I called thinking alone in the work is partly this: a refusal to offload the things that are still teaching me something.

That is the whole discipline, scaled up or down. Decide which repetitions are pure waste and let the machine have them, gladly. Decide which repetitions are secretly the training and protect them, even when keeping them costs you the very efficiency the tool was supposed to deliver. Get that judgment wrong in the cheap direction and you save time now and pay for it later in capability you no longer have. This is the slowest and least visible of the four ledgers in the real bill for AI, the one with no monthly invoice and the longest tail. Compressed or destroyed is not a property of the tool. It is a choice about which mess you decide to keep.