Germane Friction

A while ago I ran an argument against myself, on purpose, and it ended somewhere I did not expect.

I started by building the strongest case I could that leaning on AI is a mistake: that it hides its costs, decays the work, drains the worker, starves the next generation, and flattens the culture. I made it as hard as I could. Then I turned around and built the opposite case, that AI is just the next rung on a seventy-year ladder of removing effort that was never building anyone, the calculator move for knowledge work. And that one came out stronger.

But winning was not the feeling I was left with. The feeling was that I had written one essay twice. The two cases were not opponents. They were the same observation, read once as a loss and once as a gain. Both of them were about a single thing, and once I saw the thing I could not unsee it. They were both about friction, and they were each only looking at half of it.

AI is a friction-removal engine

Strip away the specifics and that is what the technology is. It collapses the distance between intending something and having it. You want a function, a paragraph, a summary, a plan, and the gap where you used to have to produce it by effort gets very small. Every honest description of what AI does, good or bad, is a description of removed friction.

So the two sides of the debate are not really disagreeing about AI. The optimist looks at the removed friction and says: that was overhead, that was waste, removing it is pure gain. The pessimist looks at the same removed friction and says: that effort was the gym, that was where the capability got built, removing it is quiet atrophy. They are pointing at the identical fact. They are assigning it opposite signs. And they are both right, about different friction, which is why neither of them can win and why the argument goes in circles forever.

Because there are two kinds of friction, and the entire confusion comes from using one word for both.

Two kinds, named plainly

I am going to borrow a word from the learning sciences, because they figured this out a long time ago and have the cleanest version of it.

When you do hard mental work, the effort splits into two kinds. Some of it is doing nothing for you. It is the noise: the badly organized interface, the boilerplate you have typed a thousand times, the syntax you already understand, the lookup, the ceremony, the part where you fight the tool instead of the problem. The researchers call this extraneous load, and their advice about it is unambiguous. Remove it. All of it. It is pure tax, and it builds nothing.

The other kind is the effort that is actually constructing something inside you. The struggle to understand, the work of building the model in your head, the fight to find the idea that was not obvious, the slow formation of judgment that only comes from having done the thing yourself. They call this germane load, germane meaning relevant, the effort that is genuinely germane to learning. And their advice about this one is the exact opposite. Protect it. This is not waste. This is the work. This is the part where you become someone who can do the thing.

So let me name the distinction the way I will use it for the rest of this. Extraneous friction is the effort that is only taxing you, and builds nothing. Germane friction is the effort that is actually building you while you spend it. Same word, friction, two completely different things wearing it.

And now the whole debate resolves, because you can finally see what each side was looking at. The case for AI is built entirely from examples where the friction was extraneous: the newcomer fighting syntax, the boilerplate server, the toll booth that kept people out. Remove that and the optimist is simply correct, it is pure gain, take it every time. The case against AI is built entirely from examples where the friction was germane: the deep work that was sharpening a senior, the apprenticeship that was forming a junior, the struggle for the sentence that was making a writer. Remove that and the pessimist is simply correct, it is atrophy, and you will not even feel it happening. They were never in disagreement. They had each grabbed one kind of friction and mistaken it for the whole.

The tool cannot tell which is which

Here is the part that makes this hard instead of just tidy, and it is the part I keep coming back to.

The AI does not know the difference. It cannot. When you ask it to do the thing, it has no way of knowing whether the effort it is sparing you was the tenth identical CRUD endpoint, pure extraneous tax, or the one piece of genuine reasoning that was about to teach you something you would have carried for a decade. It removes both with the same cheerful confidence and the same little spinner. To the engine, germane friction and extraneous friction are identical. They both look like work you would rather not do.

Which means the discrimination cannot live in the tool. It has to live in you. Every time you reach for AI, there is a question underneath the reach, usually unasked: which kind of friction am I about to remove? Is this the boilerplate, or is this the thinking? Is this the toll booth, or is this the gym? Get that question right and AI is the best thing that ever happened to your work, because it lets you delete the tax and keep the training. Get it wrong, reach for the tool on the germane friction because germane friction is hard and the tool makes it feel optional, and you will trade the version of yourself you were about to become for a slightly faster afternoon, over and over, until you are someone who ships and cannot think.

This is why the productivity studies disagree with each other. The newcomer getting fifty percent faster and the senior getting nineteen percent slower are not contradicting each other. They are the same law seen from two sides. The newcomer was drowning in extraneous friction and the tool drained the pool. The senior had already drained their pool years ago, and the only friction left in their deep work was germane, so the tool could only get in the way of it. The number is not a fact about AI. It is a fact about which friction you had left.

The skill is telling them apart in the moment

So the actual skill of this era is not prompting, and it is not knowing which model to use, and it is certainly not the binary of being pro-AI or anti-AI, both of which are just ways of refusing to do the discrimination. The skill is the discrimination itself. It is being able to feel, in the moment of reaching for the tool, whether the friction in front of you is the kind that is building you or the kind that is only taxing you, and then routing accordingly: delete the tax, keep the training, and spend the time the tax used to cost on the harder layer the tool just exposed.

That last move is the one almost everybody skips. When you let AI remove genuine extraneous friction, you get a surplus of attention back. The honest thing to do with that surplus is to spend it up a level, on the harder problem, the one the drudgery used to crowd out. The dishonest thing, the thing that turns even good AI use into slow decline, is to pocket the surplus as idleness, to let the freed attention evaporate. AI removing the right friction is only a gift if you reinvest what it gives you. Otherwise you have just bought a more comfortable way to stop growing.

I have a name for the capability this all requires, because it is the same capability I keep finding underneath everything I write about. It is complexence: the human ability to stay oriented inside something complex without fragmenting, to hold the map yourself instead of handing it to whatever offers to hold it for you. The friction discrimination is complexence pointed at AI. The tool is constantly offering to hold the map for you, to remove the effort of orienting, and some of that effort is extraneous and you should let it, and some of that effort was the only thing keeping you able to navigate at all. Knowing which is which, while a confident machine insists it does not matter, is exactly the capacity of standing in complexity without letting it, or the tool, do your thinking for you.

Where I actually stand

I am not against AI. I run agents daily, and I would not give them back. I am also not a believer in the way the loudest optimists are believers, because I have watched myself reach for the tool on germane friction precisely because it was germane, because it was hard, and felt the small private cost of having skipped the rep.

So the position is not a verdict on the technology. It is a practice, and it is a demanding one. Use AI without mercy on extraneous friction; there is an enormous amount of human effort that was never building anyone, and protecting it out of nostalgia is its own kind of foolishness. Guard germane friction like it is the thing your future self is made of, because it is. And accept that the line between them is not fixed, that it moves with who you are and what you are trying to become, that the same task is extraneous friction for the expert and germane friction for the apprentice, and that nobody, no tool and no essay, can draw the line for you. You have to draw it yourself, in the moment, again and again, which is the work, and which is also the only version of using AI well that I have ever found to be true.

That is the coin. One face says removing friction is freedom and the other says removing friction is decay, and they are the same coin, and the whole of the skill is knowing, each time you flip it, which friction you are spending. I built the case against and the case for at full strength because I had to, because you cannot see the coin until you have held both faces up to the light and found them attached. They were never two arguments. They were one, about friction, waiting for someone to say which kind.