Who Is Forming the Next Shift?

The camera was wrong, and it didn’t know it. At Lansing Delta Township, a vision system read the option code on each car so the right rear glass went in. One shift, it misread. Nothing on any screen turned red. A quality checker at the end of the line stopped at a car wearing a rear glass with an antenna the build didn’t call for.

What happened next is the part nobody wrote down. I’ll get to it.

This week, a Microsoft director of applied science surfaced in the unsealed record of the New York Times’ lawsuit against OpenAI and Microsoft. In a January 2023 memo, he called the data behind his company’s models possibly the “largest theft of labor in human history.”

He wrote labor. Not content, not data, not intellectual property, the words the lawyers will fight over. An insider looked at what went into the machine and named the people.

I have spent thirty-six years walking production lines, from Buick City to Stuttgart to Bengaluru. When someone tells me a new kind of factory has arrived, I walk it from raw material to finished good. At every station I ask two questions. Who formed the people who work here? And where does their formation go when the shift ends?

So let’s walk the AI line. It runs from wafer to token.

AI is a factory, and like every factory it runs on formed people. At three of its five stations, it takes the formation of people it paid little or nothing, crystallizes it into capital, and sells the product back against them. A plant that did that would be eating its own maintenance crew. That’s the argument.

Station one: the plant

The first station is the one a plant manager recognizes. It starts in mines and power plants, runs through chip design, and turns physical at the fab: clean rooms, process windows, yield curves, tools worth hundreds of millions held in tolerance by maintenance crews. The chips then go into data centers built by electricians, pipefitters and controls technicians, who hand the building to a thin operating crew once the lights are on.

The people here are deeply formed. A good fab technician is built over years, the way Lukas was built at Siemens Amberg, roughly 25,000 hours by age thirty. Her judgment keeps yield from collapsing on the night a chamber drifts. The electrician who wired the data center carries what she learned to the next site.

At this station the helix holds. Capital is crystallized labor; labor is capital in formation. The people who build and run it get better every year and take that capability home. Both strands compound, and a sane operator pays for both.

If AI were only this station, I would not be writing this.

Station two: training the model

Training a model is one enormous crystallization event. The raw material is formed human judgment: every well-reported article, every patient explanation, every line of code someone debugged at 2 a.m. and posted so the next person wouldn’t have to. The court filings describe one scraped dataset holding more than two million documents from a single newspaper’s site. Multiply that across every writer, teacher, coder and forum volunteer on the open web.

In a plant, crystallization is ordinary and honest. When my best operator finds a better sequence, we write it into standard work. That is capital formation. But three things hold on a plant floor. She was paid while she formed. She keeps the capability after the process absorbs it. And the process does not go to market against her.

At station two, all three break. The formers were never paid for the formation taken. The model now holds their capability in crystallized form. And the product sells into the same markets they work in. This is where one strand is harvested and cut.

Station three: the rater’s desk

A model fresh from training is raw stock. Before it ships, people finish it. They rate its answers, correct its mistakes, and write the examples it learns manners from.

This station runs on two tiers. At the bottom, annotation and content moderation flow to outsourcing centers in Africa and across the Global South, at a few dollars an hour. At the top, doctors, lawyers, engineers and coders are recruited to judge the model in their own field. One platform that matches these experts to AI labs reports paying them about $95 an hour on average.

This is the only station on the line where formation carries a price tag, and the price tells you how the buyer sees it. A physician’s twenty years is bought one rating at a time, like a gig. Nothing is capitalized in her name. The more carefully she works, the faster she trains the thing that will answer her patients without her.

Station four: the forward-deployed engineer

The fourth station is the one almost nobody talks about, and the one that should worry a plant manager most.

A model knows little about your operation. So the AI companies send forward-deployed engineers to sit with the customer’s operators, planners, schedulers and claims adjusters. They watch how the work really gets done: the workaround, the judgment call, the thing the veteran knows that no procedure holds. Then they encode it.

Any Toyota-trained reader knows the method: go to the gemba and learn from the person doing the work. The difference is where the learning lands. In a real TPS plant, the operator’s insight improves the process and raises her standing; she becomes the one who teaches the next person. At station four, the insight leaves the building, goes into a system the customer rents back, and makes the operator who supplied it less necessary.

Station two harvested millions of anonymous people at once. Station four does it firm by firm, floor by floor, with a friendly engineer and a laptop.

Station five: the token

The last station makes the finished good: the token, sold