Friction Is One of the Safeguards AI Removed

Every argument about AI in research is about the output. Is the synthesis accurate. Did the model invent a quote. Does the summary match the transcript. Reasonable questions, all of them, and there is a benchmark for each one now, a leaderboard somewhere, a slop detector. Slop detection is my favorite, because it quietly assumes we can detect the other kind.

What worries me sits somewhere else. Making the artifact used to do something to me, every time, without my noticing, and it has stopped.

The Tax You Were Glad to Lose

For years the job came with a tax. You drafted the readout. You built the deck. You wrote the synthesis one sentence at a time and argued with yourself about which quote earned the slot. Slow, dull, and I would have paid actual money to make it go away.

It is going away. And in the relief, almost nobody stopped to ask what the dull part had been doing while we hated it.

It was catching things.

The coverage gap never showed up in a meeting. It showed up at 9pm, halfway through the deck, when I realized I had four quotes about onboarding and nothing about the step the PM had actually been losing sleep over. The bad inference survived review and died later, quietly, while I was writing the sentence meant to carry it, because the sentence would not close and I finally had to admit the data did not say what I wanted. And the outlier. Nothing flagged the outlier. I found it because I happened to be the one retyping the table.

None of that was the work. It rode on top of the work, load-bearing and invisible, and I am not sure I ever saw it as its own thing until it was gone.

Two Frictions, One Box

There are two kinds of friction in this job and they feel identical from the inside, which is most of the problem.

One is toil. Formatting. Transcribing. Pasting the same finding into a third deck because three people want it in three shapes. Reconciling two spreadsheets that were never supposed to be two spreadsheets. This kind produces nothing and teaches nothing, and the vendors who want to burn it down for you are right to want that. Let them.

The other is examination. The noticing that happens while your hands are busy. It is not a step. You cannot schedule it or hand it to someone. It only happens as a side effect of going slowly enough to see what you are doing.

And the two live in the same box. That is the whole thing, really. You found the coverage gap because you were doing the tedious assembly. The examination was riding on the toil, and when the toil leaves, the examination leaves with it, free of charge, without anyone signing off on the trade.

So nobody decides to stop examining their work. They decide to skip a chore. The examination was inside the chore the entire time.

Checkpoints in a Task Costume

Once you see this you see it everywhere.

Drafting the readout was where the story either held or fell apart, usually around slide six. Building the deck was where a thin section finally admitted it had always been thin. Writing the synthesis was where some interpretation I was attached to met the evidence and lost, and I went and told someone I had been wrong about it on a Tuesday. Even sorting your notes counted. That was when the pattern surfaced. Also when its absence surfaced, which is the part people forget about notes.

Every one of these was a quality gate dressed as a task. We resented the task. We never clocked the gate.

AI does not take one of these. It takes all of them at once, and it does it without any noise, because what comes back is not a rough thing with its seams showing. It comes back looking finished.

The Failure That Hides Itself

The part of you that would catch the gap is the part going quiet, so you do not feel it. You cannot really feel it, more or less by construction. The awareness was a byproduct of the work you are no longer doing, so the awareness goes too. You miss the thing. You also miss that you are missing things.

You would think it comes for the sloppy. It comes for the careful, which is the reverse of how people talk about this.

The better you are, the more you trust your own output, the less you reread it, the fewer people in the room feel they have any standing to check it. Twenty years of being right is supposed to protect you. Here it does the opposite. The person most likely to ship something confident and hollow is the one whose confidence has never once embarrassed them.

I keep coming back to that one. The skill is intact. That is exactly the trouble, because the skill is the part telling you not to look again.

Research Was Always Just the Wrapper

Plenty of jobs are about to learn a version of this. Research learns it worse, for a specific reason.

In most work the thing and the thinking come apart. A deck gets rebuilt. A typo gets fixed three steps downstream by someone who never knows it was there. The artifact is loosely coupled to whatever produced it, and a flaw in one is not automatically a flaw in the other.

Research never had that gap. The deliverable was always just the wrapper. What you were paying for was the examination inside it, the part that earned the claim its right to be believed. The readout was never the product. It was the residue left over after the real work happened somewhere nobody was watching.

Now you can generate the wrapper with nothing in it. A synthesis that reads beautifully and sits on three sessions you half-watched on 2x. A recommendation that closes the Delta on the slide and leaves it wide open in the actual product. It looks identical to the real thing. It was always going to, because the surface was never where the rigor lived.

I am least sure how to say this next bit, so I will just say it plainly and badly. Polish used to mean something. For most of my career a clean, confident, well-built deliverable was decent evidence the work got done, because you could not make the clean confident thing without doing the work first. That is over. Polish and rigor have come apart, and I do not think they go back together. The surface of a research artifact tells you nothing now. You will have to read past it, everyone who consumes your work will have to read past it, and most of them will not bother.

You Cannot Check Your Own Work Anymore

Start with the uncomfortable part. The examination you lost was your own, and it was involuntary, which is exactly why you cannot just decide to do it again. The faculty that used to catch the gap is the same one going soft, and it cannot feel itself going. You are the last person who can be trusted to check your output, and you are also the one most likely to be doing the checking.

So the checking has to move off you and onto someone else. Two of them, really: peer review and pressure testing, done by a person who was not the author and has no stake in the claim being true.

Real peer review. The kind where someone rebuilds your conclusion from your raw data, without you in the room to narrate it. If they cannot get from the data to your claim on their own, the claim was thinner than it looked. Fresh eyes do for free what your own eyes used to do while your hands were busy, which is notice. They have not stared at this output for six hours. You have, and you stopped seeing it around hour two.

Pressure testing is the adversarial half. Someone takes your claim and argues the opposite from the same evidence, on purpose, to find out whether it survives. The questions are boring and specific. What did we not look at. Who did we never talk to. What would have to be true for this to be wrong, and did we actually check. The job is to make the thinness show up in a room, while it is still cheap.

None of this is new. Researchers have peer reviewed and pressure tested forever. What changed is that it used to be optional, a thing the careful did on top of work that was already examining itself on the way through. That second examination is gone now. So the first one has to be deliberate, and owned by someone, and that someone cannot be the person who made the thing.

It is slower. It is a worse substitute for what you lost, and I am not going to pretend it is a clean trade. You will also have to defend it to people who just watched AI make the team faster and cannot see why you want to hand some of that speed back. I do not have a tidy way to win that argument yet.

Nobody Cancelled the Bill

The tools make you faster. That part is real, and boring, and not the part I have been circling.

Speed was never free. It got paid for the whole time by the slow work that also kept you honest, and that bill did not disappear when the slow work did. It moved somewhere you cannot see it. You are still paying. You just cannot feel the money leaving.

I do not have a clean way to land this. The honest version is that the cost is real, it is already running, and the people most exposed to it are the ones least built to notice. Me included, on a bad week.

🎯 Subscribing is the one bit of friction I'll happily remove for you. Subscribe to The Voice of User.