Pre-Registered: What the Stanford AI Indicators Will Show
Published 10 June 2026. This article is a public, time-stamped pre-registration. Nothing in it will be edited after publication. Corrections, if any, will be appended and dated.
Stanford’s Digital Economy Lab is about to launch the AI Economic Indicators: a regularly updated platform tracking hiring patterns, AI usage, and consumer surplus, built on the Canaries study and the lab’s ADP payroll collaboration. Erik Brynjolfsson’s team is constructing the best labour-market instrument anyone has built for this transition.
Good. A theory that explains everything after the fact explains nothing. So before the platform goes live, here is what the Discontinuity Thesis says it will show. These predictions are specific enough to fail. If they fail, the thesis is wrong and I will say so on this page.
I encourage everyone with a competing framework to do the same. The optimist position has predictions too: hiring rebounds, displaced workers reallocate, productivity gains flow into wages. Write them down now, with dates, while the data can still embarrass you. Anyone can narrate a chart after it prints.
The mechanism, in three sentences
AI plus a human verifier beats standalone human cognitive labour on cost, and the gap is widening. No firm, industry, or nation can unilaterally refrain, because restraint is competitive suicide. And no absorption channel for displaced cognitive workers is simultaneously AI-resistant, wage-sustaining, and scalable to millions. If those three premises hold, the wage-demand circuit that sustains mass consumption severs. The Stanford dashboard will measure the severance in real time.
The predictions
Horizon: twelve months from the platform’s public launch, against its own published series.
1. The canary signal deepens. Entry-level hiring in AI-exposed occupations keeps falling, and the gap against less-exposed occupations widens rather than mean-reverting. The Canaries study found the early signal. The thesis says it was the start of a structural trend, and the optimist case requires it to have been a blip.
2. Displacement stays quiet. The mechanism is hiring freezes and attrition, with no headline layoff wave to match the hiring data. The lab’s own Labor Market Dynamism index falls: fewer hires, fewer quits, a labour market freezing from the entry rung upward. I have called this the No-Scream Principle. The dashboard will show a market going quiet, and quiet will be misread as stable.
3. The interest-rate alibi dies on the lab’s own data. The lab’s follow-up note suggested rates and timing might explain the young-worker decline. Rates have eased. If this were a rates story, entry-level hiring in exposed occupations rebounds with the cycle. It will not, because a macro variable cannot produce an effect that selects by occupation exposure and by seniority at the same time. This is the cleanest single test on the dashboard, and the lab built it against its own hedge.
4. Productivity and wages diverge. Output per worker in AI-exposed sectors rises while labour income share in those sectors stagnates or falls. The surplus is real. It accrues to capital. I have called this Ghost GDP, and the productivity series will be reported as good news while the hiring series on the same page explains where the gains came from.
5. Consumer surplus rises as the wage data worsens. The platform tracks consumer surplus from AI tools. That line goes up, and it goes up for the same reason the hiring lines go down: the tools are cheap and the workers they displace were not. You cannot pay rent with consumer surplus. Watch the two series move in opposite directions and watch one of them get labelled a benefit.
6. The narration holds regardless. Every release commentary will frame the data as transformation, augmentation, and opportunity, whatever the hiring lines show. This is the Cassandra Prison stated as a testable claim about institutional behaviour rather than an insult. The researchers are serious people producing honest data. The incentives narrate it for them.
How I am wrong
A pre-registration without failure conditions is a prophecy. Here are mine.
The thesis is falsified on this dashboard if, within twelve months of launch and with rates easing: entry-level hiring in AI-exposed occupations recovers to its pre-2024 trend; the exposure gap between occupations closes; and wage growth in exposed sectors tracks the productivity series. The thesis takes serious damage if the data shows a new occupational category absorbing displaced cognitive workers at scale and at sustained wages. Name the category, show me the payroll lines, and the three-gate test is answered.
Partial outcomes count too. If the hiring lines fall but wages in exposed sectors keep pace with productivity, the circuit is stressed rather than severing, and the thesis overclaimed its timeline. I will concede that in writing.
The invitation
Pick your framework. Abundance, augmentation, normal technological churn, rates-and-timing. Each one predicts something different on this dashboard. Publish your predictions before launch, with a date and a failure condition, and link them here. In a year we read the instrument together.
Stanford is building the scoreboard. I have just shown you my scorecard. The theories that decline to fill one in have told you what they expect to see.
The full argument is at *discontinuitythesis.com*. A standing prize remains open for genuine structural refutation.