What the Million-Dollar Tester Could Not Do

Evidence They Can’t Defend — Essay 10 of 13

Every vehicle launch has a signature failure, and on a modern car it is electrical.

Not the components. The components arrive working. What goes wrong is the connections — a harness seated proud instead of flush, an optical fibre nicked somewhere during the long violence of installing a vehicle’s insides. The part is fine. The joining of it is not. And the fault often presents nowhere near where it lives, so a customer finds it three months later in a driveway.

For the 2023 launch at Lansing Delta Township, the plan for preventing that had two parts, and I want to describe both precisely, because between them they explain the whole of this essay.

The first was training. Operators would learn the electrical work on a digital facsimile of the vehicle, in the trim shop — clicking on parts, clicking on points in an image of the car, clicking tools, completing sequences on a screen.

I thought then and think now that this is a classic formation error. There is no substitute for a master showing, instructing, observing, and certifying. Toyota has a word for the place where that happens: the dojo, a training area off the line where a new operator practises the actual work under a seasoned hand until someone competent signs off that they can do it to standard. Then they take a station. We had a screen with clickable parts.

I have watched that error take the same shape for thirty-six years. Somebody skips the formation, buys an instrument to catch what the formation would have prevented, and then defends the instrument long after it stops working. What follows is the cleanest example I have, and it cost a million dollars.

The second part was inspection. To stop cars leaving the trim shop with electrical faults, the plant installed about a million dollars of equipment that plugged into the vehicle’s electronics and verified circuit completion and function.

There was one problem, and it was mine. There was not enough cycle time for the test and not enough footprint to put it anywhere except at the end of the trim shop — which is to say, at my waypoint. I had thirty cars between me and chassis. Every blip on that tester ate into the bank, and if the bank emptied, chassis starved and I became Herbie: the constraint the whole plant is waiting on.

So watch what the million dollars actually bought. False alarms. Cycle-time problems from the equipment and from the inspectors operating it. And diagnosis time — because when the tester flagged a circuit, somebody still had to find out what was wrong with it, and finding out is slow.

Every one of those costs pointed back to the same place. Not to the tester. To the training that never happened: no dojo, no verified electrical skills, no repetitions on this particular vehicle under someone who could confirm the work was right.

And here is the ending, which is why I am telling you this in an essay about algorithms. Within about six months of ramp-up, the operators had accumulated enough repetitions to climb the learning curve on their own — the hard way, on live cars, without a dojo. The tester’s catch rate fell to almost nothing. It had become expensive and useless at the same time.

We disabled it. For good.

A million dollars of inspection technology, deployed as a substitute for formation, switched off six months later because the humans had finally learned what nobody had taught them.

That’s the argument. Every extraction mechanism in this series required a human decision. The gig economy automated the deciding — and what it automates, over and over, is the same swap I watched at LDT: a measurement system standing in for a formation system, priced as if it were cheaper, and defended long after it stops working.

Algorithmic wage suppression

Traditional wage suppression works through institutional mechanisms: weakened unions, stagnant minimums, industry norms. It operates over years.

Algorithmic wage suppression operates in real time.

Ride-share platforms can adjust the per-mile rate a driver receives minute by minute. If the algorithm detects that drivers in a market have limited alternatives, it can lower rates without any human deciding to suppress a wage. The suppression sits inside the optimisation function.

A quarter of gig workers earn less than the state or federal minimum wage once all hours, including waiting time, are counted. Because they are classified as independent contractors, minimum wage law does not apply. The misclassification of the previous essay, industrialised.

The Amazon model

Amazon warehouses are where digital extraction meets physical labour. Workers are tracked by algorithms measuring “time off task” — any period not actively scanning, picking, or moving product.

A 2023 OSHA investigation found Amazon warehouse workers suffering serious injuries at more than twice the rate of non-Amazon warehouses.

This is the logical endpoint of Frederick Taylor’s scientific management. Taylor wanted to separate thinking from doing — management determines the one best way, workers execute without deviation. The algorithm completes it: determines the method in real time, monitors compliance continuously, terminates automatically.

Taylor’s stopwatch is now continuous sensor data. His foremen are automated tracking. And his suppression of worker intelligence is perfected by a system that does not need to suppress intelligence, because it has no input channel for it.

That last clause is the one I would underline. My tester had no input channel either. It could tell me a circuit had failed. It could not tell me the operator had never been taught to seat the connector, which was the only fact that mattered.

The extraction arithmetic

Uber went public in 2019