Episode Details

Evidence: David Autor and Neil Thompson, “Expertise,” NBER Working Paper 33941, 2025.

[https://www.nber.org/papers/w33941](https://www.nber.org/papers/w33941)

Popularized as: Seb Murray, “A new look at how automation changes the value of labor,” MIT Sloan Ideas Made to Matter, August 18, 2025.

[https://mitsloan.mit.edu/ideas-made-to-matter/a-new-look-how-automation-changes-value-labor](https://mitsloan.mit.edu/ideas-made-to-matter/a-new-look-how-automation-changes-value-labor)

To drive a black cab in London you had to learn the Knowledge.

Twenty-five thousand streets inside a six-mile radius of Charing Cross. Twenty thousand landmarks. Not the map — the routes. An examiner names two points and you speak the run aloud, turn by turn, from memory, in real time, while he tries to break you. Three to four years, most of them. Most of them failed.

Neuroscientists at University College London went looking at the brains of the ones who passed. The posterior hippocampus — the structure that holds spatial memory — was measurably larger than in matched controls. Not larger in the men who were going to pass. Larger in proportion to the years spent learning.

The organ changed shape.

That is what formation is. Not a certificate. A physical alteration in a human being, produced by time under load, which cannot be produced any other way and cannot be produced faster.

Then GPS arrived.

Neil Thompson, who studies this at MIT, describes what followed in a sentence: the deep local street knowledge had been the real differentiator, and once it was automated the service commoditized. Lower wages. Many more drivers.

The Knowledge did not depreciate. Nobody forgot it. It was repriced to zero, and then it stopped being made.

What the paper found

Autor and Thompson set out to solve a puzzle. Some occupations were heavily exposed to automation and never suffered the wage collapse the models predicted. Some got paid more.

Their instrument is elegant. Skip job titles, skip surveys — read the language used to describe the actual tasks. Specialized vocabulary marks specialized work. Run that across three hundred–plus occupations, 1980 to 2018.

The result splits along one seam: which end of the job got taken.

Automate the simple end and the remaining work demands more expertise. Bookkeepers and accounting clerks lost a third of their employment across those decades — and their real hourly wages rose nearly forty percent.

Automate the expert end and the job opens up. Inventory clerks faced the same computerization, but it took their most expert tasks. Employment more than doubled. Real wages fell thirteen percent.

Same technology. Opposite sign. The variable is not how much automation arrives. It is what the automation eats.

Good paper. Real finding. Now what it cannot see.

Watch the verb

Here is the mechanism in the paper’s own framing: when the simple tasks go, the surviving work pays more because fewer people are qualified.

Are qualified.

Not become qualified.

Read the whole apparatus and you will not find a channel by which anyone enters the pool. The pool narrows. The survivors get repriced. Where the survivors came from is not asked, because the instrument cannot ask it.

This is not an oversight by two careful economists. It is a property of the measurement. Expertise here is an attribute of a task, inferred from vocabulary, measured across a moment. It is a stock. There is no flow term. Nothing in the framework can detect accumulation — only the residue of accumulation, after it has precipitated out into a wage.

The paper sees expertise everywhere it has finis