Economists are arguing about the wrong variable

The economics of automation is a well-developed literature and it has a clear answer about the long run: labour reallocates, new tasks emerge, aggregate employment recovers. That answer is well-evidenced across two centuries, and I believe it.

It is also being deployed to answer a question it does not address. The long-run equilibrium tells you where the system settles. It says nothing about how long the settling takes, and the historical record on that is much less comforting.

The mechanisation analogy, read properly

Agricultural mechanisation is the standard reference. Employment in agriculture fell from a majority of the workforce to a few per cent, and the displaced labour was absorbed. True — over roughly a century, across three generations, with most individual workers never transitioning at all. Their grandchildren did.

That is a successful adjustment at the level of the economy and a catastrophic one at the level of a life. When people cite this analogy as reassurance, they are citing an aggregate statistic to answer a distributional question.

Entry-level contraction is the leading indicator

The data I find most concerning is not aggregate employment, which is noisy and lagging. It is the composition. In occupations with high task-automation exposure, junior postings are contracting faster than senior ones, and the effect is strongest where the output is text or code.

This matters more than it first appears. The tasks used to train junior practitioners are the most automatable tasks — that is nearly a definition. If those tasks disappear, the pipeline that produces senior practitioners closes, and the effect is invisible for a decade and then irreversible.

You do not notice a closed pipeline until you need the people it was supposed to have produced.

Why the productivity numbers are disappointing

Measured productivity gains have been smaller than adoption would predict, and the verification argument explains most of it: gains are bounded by the cost of checking the output, and in strict-correctness domains that bound is close to zero.

I want to add a distributional point. Verification labour is real work that is expanding, and it is systematically worse work — more monitoring, less creation, less legible as skill. "Productivity is flat" and "the work has gotten worse" are entirely compatible, and only one of them shows up in the statistics.

What I would want measured

• Transition rates, not employment levels — how many displaced workers reach comparable work, and how long it takes.
• Entry-level composition by exposure, tracked continuously rather than in retrospective studies.
• Verification labour as a distinct category. It is currently invisible in occupational statistics.
• Geographic concentration of both displacement and the new work, which historically have not been in the same places.

None of this requires taking a position on AGI timelines. These effects are measurable now, at current capability, and they will be worse before any threshold in the Index is crossed.