Issue · July 21, 2026

The measurement argument

More than four hundred economists signed a statement saying AI could reshape the economy faster than the Industrial Revolution did, and that nobody can currently measure it. Both the people who want to legislate now and the people who want to wait are arguing from that same admission.

What was signed

On July 13 Fortune reported that more than two hundred economists had put their names to a statement organised out of Stanford’s Digital Economy Lab by Erik Brynjolfsson, Anton Korinek, Ajay Agrawal and Tom Cunningham. The list has since passed four hundred. Sixteen signatories are Nobel laureates. The chief economists of OpenAI and Anthropic signed, along with Joseph Stiglitz, David Autor, Diane Coyle, Jason Furman and Eric Schmidt.

The text itself is a page long. AI “may become radically more powerful over the next 10 years” and could drive a transformation “larger than the Industrial Revolution, but unfolding over a vastly shorter time frame”. It asks for work on “the incentives, guardrails, and institutions needed to steer AI”.

What carried the story was the admission underneath it. Fortune’s headline quoted the phrase the economists were using about themselves: driving in the fog. A profession that publishes quarterly forecasts said, collectively, that it cannot currently see the thing it is forecasting.

The case for moving before the data

Michael Spence, the NYU Nobel laureate, called for an all hands on deck response. The argument is about lead times rather than confidence. Wage insurance, retraining systems and the tax treatment of capital against labour take years to legislate and longer to work. If disruption arrives on anything like the schedule the statement sketches, a policy that waits for clean statistics arrives after the population it was meant to protect has already moved on.

Daron Acemoglu signed, which is the detail worth sitting with. He has spent years arguing that the measured gains from AI are modest and that much of the surrounding discourse is inflated. His name on a document about urgent preparation suggests the uncertainty is running in both directions, and that scepticism about today’s numbers is not the same as confidence about the next decade.

The case that the evidence cannot carry it

A day after the Fortune piece, Torsten Slok, chief economist at Apollo, published a note on how badly the underlying measurements disagree. Five frameworks score occupations for AI exposure on a scale from zero to one: Massenkoff and McCrory in 2026, Felten and co-authors in 2021, Eisfeldt in 2023, Eloundou in 2024, Tomlinson in 2025. On hairdressers and dancers they agree almost exactly. On tax preparers, telemarketers and mathematicians, which are the jobs the entire argument is about, they diverge sharply. Slok’s summary is that the jobs most likely to be called at risk are the ones we understand least.

He had made the companion point a week earlier. Between the first quarters of 2023 and 2026 the Magnificent Seven lifted profit margins from roughly 15% to 25%, while the other 493 firms in the S&P 500 sat flat near 10%. He cites an MIT study finding that 5% of companies saw meaningful return from their generative AI pilots. His conclusion is that the gains so far belong to the companies selling the technology, and that a painful repricing follows if the rest of the economy keeps not finding them.

Two refusals

Both landed on 15 July, from opposite directions. Ben Goertzel gave his reasons for not signing while accepting the premise outright. His objection is to the phrase about steering AI to complement humans, which he reads as the Acemoglu school arriving through a side door. Nobody can predict what a given system will do, so on his account complementarity has to fall out of the traffic between billions of people and billions of agents rather than out of a policy brief. He wants many architectures and value systems running at once.

Noah Smith declined from inside economics, on procedure. The statement asks for a signature without naming a policy, which he treats as endorsing recommendations he has not been shown. He presses the forecasting point harder than Slok: nobody knows at the moment of invention whether a technology will complement labour or replace it, and the employment data so far shows no AI job destruction to explain. He adds that Acemoglu’s own predictions about AI and labour have not aged well, which sits oddly beside Acemoglu’s signature. Smith would sign a version that said what acting now means.

Our read

The statement is being read as a warning about AI. It is closer to a warning about the instruments. Strip out the forecasting and what four hundred economists actually signed is a claim about how to behave when you cannot see: build the capacity to respond before you can prove you need it. That is a position about risk, not a finding, and it should be argued on those terms rather than waved through because sixteen Nobel laureates are on the list.

Slok has the better of the narrow point. An exposure literature where five methods agree about dancers and split on tax preparers cannot carry the weight of a labour statute, and anyone citing a specific displacement number this year is citing a choice of framework. But his objection lands on the forecasts, not on the request. Retraining systems and wage insurance take years to build whether or not the disruption shows up, and the cost of standing them up early is mostly wasted money, while the cost of standing them up late is measured in people.

What we would watch is who pays for the instruments. The statement asks for research and institutions in the same breath, and the research half is cheap, uncontroversial, and the only part that would settle any of this. If a year from now the Bureau of Labor Statistics still cannot say which occupations are being automated, the fog will have become a choice.

The measurement argument

What was signed

The case for moving before the data

The case that the evidence cannot carry it

Two refusals

Sources

One email whenever an issue ships