The Case Against Letting AI Think For You

This is one side of a three-part argument. I built the opposite case too, and it came out stronger. The resolution, where I actually stand, is here. But I want to make this side properly first, because it deserves better than the version people usually wave away.

In early 2025, sixteen experienced developers sat down to do real work in repositories they had maintained for years. For half their tasks they were allowed AI tools. For half they were not. Before they started, they expected the AI to make them about a quarter faster. After they finished, they were sure it had, by about a fifth. Then the researchers measured what actually happened. With the AI, they were nineteen percent slower.

That study, run by METR, is the most uncomfortable single fact in this whole debate, and not because of the slowdown. The slowdown is arguable; it was a specific cohort doing deep work in code they already knew cold, which is the exact case where AI helps least. What should bother you is the gap between the nineteen percent slower they measured and the twenty percent faster the developers felt. These are good engineers. They could not feel the truth about their own productivity. The tool did not just change their speed. It changed their ability to perceive their speed.

That is the shape of the entire case against leaning on AI. It does not break your work in front of you, where you would notice and stop. It moves the cost somewhere you cannot see.

It moves the cost to later

Start with the artifact, because it is the easiest place to watch the cost migrate.

GitClear looked at more than two hundred million lines of code written between 2020 and 2024, the window where AI assistants went from novelty to default. The trend lines all point the same way. Copy-pasted code climbed. Refactoring, the work of taking something that works and making it cleaner so the next person can stand on it, fell by roughly sixty percent. Duplicated blocks multiplied. Churn, the share of code rewritten within two weeks of being committed, went up: more code that was wrong on arrival.

None of this is the AI writing garbage. The individual suggestions are usually fine, even good. The problem is what gets skipped. Designing a reusable abstraction is slow and effortful, and an autocomplete that will hand you a working copy of the nearby thing makes skipping that work feel free. So you skip it. Then you skip it again. Each skip is invisible and cheap. The bill arrives months later, in a codebase that nobody refactored because refactoring stopped being the path of least resistance, and now every change touches eight near-duplicates of the same idea.

This is technical debt, but worse, because technical debt is usually a decision. Someone says, out loud, we will do this the quick way now and pay for it later. The AI version is debt you take on without ever deciding to. The cost was simply moved downstream, past the point where you were paying attention.

It moves the cost to you

Now the harder migration, because this one comes out of the person.

There is a growing pile of research, across coding and writing and reasoning, on what happens to a mind that offloads its thinking to a machine. The findings are consistent and they are not subtle. Heavy reliance correlates with weaker critical thinking. People given an AI assistant give up sooner when it is taken away and perform worse on their own afterward. Researchers have a clinical word for it: deskilling. The capability you stop using does not hold steady. It drains.

The mechanism is not mysterious, and it is worth being precise about, because it is the hinge of the whole argument. The struggle you are tempted to delegate is, very often, the exact struggle that was building you. The half hour of being stuck on a problem is not a tax on the learning. In the learning sciences it has a name, desirable difficulty, and it is the part of the work where the understanding actually forms. The fight to find the non-obvious sentence is where you become a writer. Hand the fight to a machine and you get the sentence. You do not get to become the writer.

This is the trade that feels like a gift and is not. Every time you let the tool do the part that was hard, you get the output and you skip the rep. One skipped rep is nothing. A thousand skipped reps is a different person, one who can no longer do the thing they used to be able to do, and who may not even remember that they used to.

It moves the cost to the next person

Watch the cost leave the individual and land on a whole generation.

Junior engineers are not learning the way they used to, because the work juniors learned from is exactly the work that AI now absorbs. The boring function, the careful debugging, the slow apprenticeship of seeing what your choices cost two weeks later: that was never just output. It was how an engineer got made. When the tool does it instead, the company gets the output and the human gets none of the formation.

The hiring numbers have started to follow. Entry-level developer postings have fallen hard since 2022. Senior roles have held or grown. Read that forward and the picture is grim in a slow way: if you stop making juniors, you stop making the mid-levels they would have become, and you stop making the seniors after that. The pipeline that produces the people who can actually judge AI's output is being quietly defunded by the same tool that made it look optional. Learning cannot be outsourced. But hiring is trying to outsource it, and the bill comes due in about a decade, on someone else's watch.

It moves the cost to the culture

The last migration is the widest and the easiest to miss, because you cannot see it from inside one person's work at all.

Large language models converge. Independently trained models from different labs on different continents produce strikingly similar output, measured in one study at around eighty percent overlap. When people write with AI, their individual pieces often get more polished and the whole set gets less diverse. Researchers call the attractor an artificial hivemind: not one voice silencing the others, but a million voices drifting toward the same mean because they are all leaning on tools that pull in the same direction.

Run that across a culture for a decade. Not censorship, nothing so dramatic. Just a slow flattening, a narrowing of the range of things that get said, because the frictionless path always runs through the middle of the distribution. The weird sentence, the idiosyncratic structure, the idea that sounds wrong until it doesn't, these are precisely the things a model trained on the average will smooth away, gently, every time, for free.

The crack in my own case

Here is where an honest version has to turn on itself.

Every harm I just laid out is a harm of how the tool is used, not a property of the tool. Slower in deep work, fine: it is faster in shallow work. Code decay: that is a discipline that AI happens to make easy to drop, not one it forbids. Deskilling: it offloads the struggle that builds you and also the struggle that just wastes you, and I have not yet told you how to tell those apart. Homogenization drops back to baseline the moment people use the tools with any deliberateness at all.

Notice what that means. I built this whole case on a single move: the cost gets hidden and paid later, by you, by the next person, by the culture. But I never said which costs were worth paying. I treated all friction as sacred, and that is not true, and the strongest reply to everything here is to point at the mountain of human effort that was never building anyone and ask why I am so eager to protect it.

That reply is real. I made it in full, as hard as I could, and it came out stronger than this did. Both of these essays are one claim about the same thing, read with opposite signs, and the thing is friction. The question is not whether AI removes it. It is which kind it is removing, and I had to argue both sides at full strength before I could see that the fight was never really about AI at all.