The mean and the tail

Two readings of the same economy landed within hours of each other this weekend. At Jackson Hole, global central bankers spent their annual retreat gaming out a dystopian AI future, per Reuters. The same morning, Alex Tabarrok published the most direct evidence anyone has on the question — Census Bureau surveys of hundreds of thousands of firms — under the headline So Far, So Good.

Both are right. Not "the truth is somewhere in between" — both, fully, at the same time. The reassurance is a statement about the mean. The dread is a statement about the tail. And in the Census numbers the mean is flat while the tail is the one part that's moving.

The mean first. When the Bureau started asking in late 2023, 3.7% of firms said they'd used AI in the prior two weeks to produce goods and services; by late 2025 it was about 10%, and 18% under a broadened question. Twice it has asked firms directly how AI affected total employment. In early 2024: 2.8% said it increased, 2.6% decreased, 94.6% no change. Two years later: 2.3% up, 2.0% down, 95.7% no change. Adoption nearly tripled; the employment effect didn't budge.

Now the tail, from the same release. Among firms where AI has taken over employee tasks, the share reporting it took over "a large number" of them tripled, from 2.4% to 7.1%. "A moderate number" went from 13% to 22%. That group is small — roughly a tenth of adopters, who are themselves about a fifth of firms — which is exactly why it doesn't show up in the topline. A tripling from a small base inside a subpopulation is what every diffusion curve looks like in the years before it becomes a mean effect. The average is calm. The derivative isn't.

The average is calm. The derivative isn't.

The skeptic's version: central bankers war-gaming science fiction at a mountain retreat is attention following vibes, not data, and policy anchored to imagined futures has a poor record. Concede most of that — monetary policy should key off measured effects, and the measured effect here is a rounding error. But stability institutions aren't paid to price the mode of the distribution. They're paid to price the tail, and to price it before it prices itself. The same weekend, the Financial Stability Board's watchdog named AI-driven cyber risk the top concern for global financial stability. That's not a panic. That's the job — the survey measures what happened; the stability crowd is accountable for what could.

So read the weekend's split screen as a division of labor, not a disagreement. "So far, so good" is the mean talking, and the mean deserves to be heard — it has two years of consistent, large-sample data behind it. The tail tripled. Both numbers were in the same release.