UNDERTOW 015: Tacit Harvest

AI is learning to do the work no one could ever explain, by recording the people who do it, then setting them aside.

In June 2026 a cleaning company in Singapore put robots to work in office buildings and kept the human cleaners on. The humans wore the equipment. Sensors on the arms and the back, cameras at eye level, a rig that recorded how a person decides which surface to clear first, how hard to press, when to stop. The robots cleaned. The people generated the data that would let the robots clean without them.

The arrangement was reported as a robotics story. It is a harvest story.

The same week the South Korean trade ministry described a 48-billion-won program to convert the knowledge of master workers into machine-readable datasets across thirty manufacturing processes. The same month Beijing directed 10,000 humanoids into real jobs by year’s end, on the stated logic that a machine learns the work only by doing the work. Three governments, three industries, one motion: put a recorder on the part of a person that cannot be written down, and run it until the writing is no longer needed.

The part that cannot be written down has a name. The chemist Michael Polanyi called it tacit knowledge and reduced it to a sentence: we can know more than we can tell. The welder knows when the puddle is right and cannot give you the rule. The nurse knows which patient is about to crash and charts nothing that would explain how. The cleaner knows the building. For seventy years this was the floor under manual labor, the reason automation kept arriving and kept stalling. You cannot program a judgment its owner cannot state. Rule-based machines hit that wall and stopped at it, every time, for decades.

Machine learning did not climb the wall. It went around. It does not ask the welder for the rule, because the welder never had one. It watches the welder weld, ten thousand times, and learns the judgment the way an apprentice learns it, from the doing and not the telling. This is the thing to hold still and look at. The skill was never cracked. It was never reduced to instructions. It was apprenticed to, recorded, and copied. An apprentice that never forgets, never tires, never ages, and duplicates itself at no cost does not need the master once the recording is good.

There is one prior case of irreplaceable living material taken once, made permanent, and copied without end. It is not from labor. It is from medicine.

In 1951 a hospital in Baltimore removed a sample of cancer cells from a woman named Henrietta Lacks. No one asked her. The cells did what almost no human cells do: they kept dividing outside the body, and they are dividing now, in laboratories on every continent, seventy-five years on. They built the polio vaccine on them. They built an industry on them. Lacks died that year, poor, and her family learned what had been taken decades after the fact. Her family is still finding out. They are in court now, winning settlements from companies that built fortunes on those cells — Thermo Fisher, then Novartis this past February, in suits that call the cells what the family always has: stolen. The recourse is real, and it is generations late. Henrietta got none of it.

Henrietta Lacks was a Black woman, and what was done to her belongs to a specific history of Black bodies used by American medicine without consent. A welder is not in that history, and does not die of what is taken from him. The harms are not the same. The shape is. Something that normally dies with the individual. Taken without asking. Made to persist past its source. Copied without limit. The source severed from the result. The value gathering downstream, with the one who took it. That is the cell line. It is also the rig on the cleaner’s back. Biologists have a word for a sample made to outlive its donor and reproduce on its own. They say the line has been immortalized. They call the product cultured.

This is what I am calling Tacit Harvest: the recording of un-writable human skill, once, for permanent copy, from a worker who is then set aside. The Korean ministry is building a severance dataset. The Singapore cleaner is being cultured. Three names for one motion. Keep the one you cannot shake.

Start with what has already happened, because most of it has. Rule-based automation spent forty years failing to climb the wall. Capture cleared it in a few. In Chinese logistics this spring a humanoid, built on exactly this kind of recorded work, reached as much as 85 percent of human efficiency at a task a decade of programming could not touch. Eighty-five percent is not the finish line. It is the stretch of the curve that used to be impossible, crossed in months. The rest closes the way every machine-learning gap closes. More recordings.

The machine will close that last fifteen percent eventually. The worker does not get to wait for it. Most people read the gap as his security, the margin that still needs him; it works the other way. He is not replaced when the copy becomes perfect. He is replaced when it becomes good-enough at a tenth of his cost, and good-enough is already here. A perfect copy would still need him for the edge cases. An eighty-five-percent copy that never tires and never bills overtime means no one pays for the last fifteen percent again. He is not undone by the machine that does the job better. He is undone by the one that does it worse and cheaper, and the market files that under progress. The line never had to reach a hundred. It cleared at eighty-five, months ago, while everyone was still watching the curve climb.

The obvious objection arrives here, and it is correct as far as it goes. None of this is new. Knowledge has always moved from the skilled to the unskilled by watching. The apprentice stood at the master’s elbow for years and absorbed what the master could not say. This is how the trade survived. Capture, the objection goes, is only apprenticeship at scale.

It is apprenticeship with three things removed. The old apprenticeship took years; the recording takes a shift. The old apprenticeship produced one successor; the recording produces an unlimited number, at no cost, forever. And the old apprentice did not replace the master. He stood next to him, and the master kept working, and the skill still had somewhere to live that was a person. The new apprentice replaces the master the moment the file is good. The welder who taught it is walked to the door. He taught it well, which is why he is no longer needed. The behavior is old. The consent, the scale, and the discard are new, and they are the whole of it.

Then the recording detaches. Once the skill is on file it stops belonging to anyone. It can be bought, licensed, versioned, shipped to a factory that never met the worker it came from. The skill that lived in a body now lives in a corpus, and a corpus is property.

This is where the second objection lands, and it is the better one: nothing was stolen. The machine was not handed the welder’s rules, because there were no rules. It learned the way the welder learned, from doing. This is true, and it does not soften the case. It sharpens it. The claim was never that a secret was extracted. The claim is that the apprenticeship itself, the oldest way one person hands skill to another, has been turned into an industrial input and the source made disposable. Nothing was stolen because nothing needed to be. He was recorded, and recording was enough.

Which leaves the worker one piece of leverage, and it is not his wage. The wage is already gone; the body costs less every quarter. Morgan Stanley’s survey found 92 percent of Chinese industrial buyers won’t adopt at scale until a humanoid drops under roughly 28,000 dollars, and the market is sprinting at that line. Below it, the human’s labor is the expensive variable and the camera is the only thing he still controls. The first labor fights of this era will not be over pay. They will be over the right to refuse the rig.

There is a kinder reading, and it is the hardest to answer. South Korea is not discarding its masters. It is saving them. The knowledge of a sixty-year-old machinist dies with him if no one records it; the dataset is preservation, and a culture that keeps the skill has lost less than one that loses the skill and the man. The philosopher Bernard Stiegler had a word for a thing that is medicine and poison in one dose: pharmakon. The dataset is one. It saves the knowledge and ends the line that produced it, and it does both with the same recording. The master is preserved. And yes — the ministry says the dataset will train the next generation, and it will. But a worker trained from a file learns to run the master’s recorded judgment, not to grow his own; he inherits the output and skips the years that made it. That produces a minder of the recording, not an heir to the craft. The skill is preserved. The path to owning it closes quietly behind him. What you are left with is a fossil, perfect and finished, of the living thing it replaced.

So the question stops being who can do the work. The machines can nearly do the work. The question is who is worth anything once the work is recordable, and the answer is the strangest part.

Follow the price down and something inverts at the bottom. When every recorded skill is cheap, the only expensive skill is the one not yet recorded. Value stops attaching to the work and attaches to the absence of a file. The master welder’s worth is no longer that he can weld; machines can nearly weld. His worth is that no one has filmed him yet. He is valuable in the exact measure that he is unharvested, and that value holds until the first good recording and not one shift longer.

This is the thing the cell line already knew. Henrietta Lacks was not valuable to the people who took her cells because she was a person. She was valuable because the cells were rare, and alive, and not yet cultured. The moment they were cultured she was finished, and the line went on without her, dividing in rooms she never entered, on continents she never saw. The cells did not need her. They had her.

I am paid for a verdict I reach before I can defend it. Two near-identical things land in front of me — two prototypes of an interactive flow, two cuts of a film, two headlines that say the same thing and don’t — and one is alive and one is embalmed, and I know which in less time than it takes to say so, the knowing landing under the sternum a half-second before the first reason arrives.

Walter Murch, who cut Apocalypse Now and The English Patient, found the words I never could: he cuts on the blink, the instant a thought finishes behind the eyes, and admits the elaborate rules he is famous for are a fiction built to carry a feeling that had already arrived. The reasons I give come the same way — afterward, dressed as rigor, numbers I slide around a deck until they total the verdict I already had. I have done his exact thing for a living, and the only difference is that his stayed in the edit bay and mine never stayed in one medium: the same call in a film, an interface, a sentence, the whole way a company walks into a room.

I work now beside the machine built to record that half-second. It does the making already, faster than any of us, and it is learning the choosing the only way the choosing can be learned — by watching people like me choose, ten thousand times, and never once being told the rule, because there was never a rule to tell. The recording is reaching me. I used to think my taste was the most me thing about me, the part no one could copy because I could not explain it even to myself. I had it backward. The thing I cannot explain is the easiest thing in the world to take — it was never going to be argued out of me, only recorded off me. I thought not being able to say it was my moat. It was the spec. I am the welder, and no one has filmed me yet.

You will want to file this under factory work. Welders, cleaners, machinists, the manual trades, somebody else’s body in somebody else’s plant. That was the safe version a decade ago, when the moat was the body and the body belonged to someone you are not. And it is not somebody else’s country anymore. DoorDash has paid delivery drivers to film themselves doing chores at home; data firms are paying people to strap a camera to their foreheads and film their own hands. The moat is gone now, and it was never only theirs. Whatever you know that you cannot say, the judgment you would struggle to write into a procedure, the thing that makes you worth more than your job description, is the part that cannot be automated until it is recorded. It is also the part now worth recording. You are not outside this. You are earlier in it.

And I should say where I stand, because it is not where the welder stands. The same machine reaches for my judgment and for his hands. But he loses the work that fed his family, and I lose an edge and keep the title that let me see it coming.

The shape is shared. The cost is not.

The cleaner was not asked. Lacks was not asked. You will not be asked, because consent was never part of the design. It was the friction the design exists to remove. What outlives you is the one thing about you that could never be written down, copied now without end, in rooms you will never enter. The cells did not need her. The file will not need you. No one asked. No one will. The recording is enough.

---

---

What To Brief From This

The capture brief. For any org whose edge sits in a few people who know more than they can say. The decision is no longer whether your experts’ judgment gets recorded. It is whether you record it or a vendor does. Once that judgment becomes a file it becomes property, and property has an owner; the only question is whether the owner is you. Brief: audit where your irreplaceable tacit knowledge lives, and decide — deliberately, this year — to own the capture pipeline in the parts of the business you cannot afford to rent.

The positioning brief. For brands and firms that sell expertise or taste. When every recorded skill is cheap, value moves to the skill not yet on file. The premium is no longer “we can do the work.” It is “we are making the call live, in the room, before it has been reduced to a recording.” Brief: stop selling the deliverable (recordable, soon cheap) and start selling the judgment-in-motion — the currently-unharvested faculty that the recording cannot reach until after you have used it.

The consent brief. For people leaders and ops. The first labor fights of this era are over the right to refuse the rig, not over pay. How you record your people will become a trust and retention issue before it becomes a legal one. Brief: get ahead of consent now. Treat the terms under which your workforce is captured as a stated, negotiated thing — not the friction you quietly engineer away — because the firms that strip it without asking will be the ones the talent leaves first.

The arguable bet. The forward position, for builders. The winning seat in five years is not the unrecorded worker. It is the firm that owns the cleanest capture pipeline in its category and licenses the judgment back. The labor fight and the data land-grab are the same fight, and most orgs are staffed for exactly one of them. Brief: if you are building, the live question is whether you own the recording or rent it — and whether you are organized to fight on both fronts at once, because they are one front.