Yale Budget Lab says AI's effect on aggregate employment is statistically indistinguishable from zero. Stanford payroll data shows software developers aged twenty-two to twenty-five down twenty percent while workers over thirty are up. The aggregate number is the disguise.
The Yale Budget Lab reported in February 2026 that the aggregate macro employment effect of artificial intelligence is statistically indistinguishable from zero. Aggregate hiring continues. Aggregate wages keep rising. The recession that AI was supposed to cause has not arrived in the headline data. By the measurement most economists watch, nothing is happening.
The Stanford Digital Economy Lab, working with the same ADP payroll dataset that covers roughly twenty-five million American workers, reports something different. Software developers aged twenty-two to twenty-five are down approximately twenty percent from their late-2022 peak. Workers in the same high-AI-exposure occupations who are over thirty are up six to thirteen percent over the same window. The split is specific to high-exposure jobs. Low-exposure occupations show no age divergence at all.
Both findings are correct. They describe the same labor market. The aggregate zero is the average of a redistribution.
The Mechanism
J. Scott Davis at the Dallas Fed published a paper in February 2026 that explains the redistribution. He split occupations by experience premium — the wage gap between a twentieth-percentile and an eightieth-percentile worker in the same job. Lawyers, credit analysts, and physicians show experience premiums above one hundred percent. Fast food workers, ticket agents, and warehouse pickers show premiums under ten percent.
When AI exposure is added to high-experience-premium occupations, wages rise. When the same exposure is added to low-experience-premium occupations, wages fall. The two effects do not cancel because they redistribute earnings rather than destroy or create them. A lawyer using AI bills the same hours at higher rates. A ticket agent competing with a kiosk earns less per shift.
Davis's explanation is that AI replicates codified knowledge well and tacit knowledge poorly. Codified knowledge can be written down: legal precedents, accounting rules, the steps to refund a flight. Tacit knowledge is built through practice and cannot be downloaded: knowing which client will object to which clause, which financial statement is hiding what, which surgical complication signals which underlying problem.
Where the experience premium is large, tacit knowledge is most of what the job is. AI augments the codified portion and the senior worker captures the gain. Where the premium is small, codified knowledge is most of the job. AI substitutes for the worker and the wage falls.
The Junior Tax
The age split inside high-exposure occupations is the secondary effect of the same mechanism. Tacit knowledge has to be built. Building it required junior workers to perform the codified work that AI now performs more cheaply. The bottom rung is the apprenticeship layer, and apprenticeship was always a money-losing investment that paid back in tacit knowledge years later. With AI handling the codified portion, the investment no longer pencils.
The data reflects the arithmetic. Computer science graduates in the United States face an unemployment rate of approximately 6.1 percent, well above the all-graduate average. Software developer job postings rose substantially in early 2026, but a widely cited industry analysis estimates that hiring conversion fell sharply over the same window — postings well above pre-pandemic levels while actual hires lagged. Companies are screening more and committing less. The pattern is consistent with employers waiting to see whether AI tools will reduce the number of junior hires they actually need.
Forrester predicted in October 2025 that fifty percent of AI-attributed layoffs would be reversed by 2027. The reversal does not mean the workers come back. Forrester's mechanism is that companies rehire at lower wages, often offshore or in lower salary bands. Fifty-five percent of employers in the same study reported regretting their AI-driven workforce reductions. They regret the execution, not the direction. The jobs that come back come back cheaper.
Why the Aggregate Hides It
Macro employment numbers cannot capture this pattern because they sum across occupations and ages. A twenty-three-year-old software developer leaving the labor force is one job lost. A thirty-eight-year-old software architect taking a larger bonus is one wage gain. Headline employment treats them as offsetting. The redistribution is invisible until the data is sliced by both age and AI exposure together, which is exactly the slice the Stanford team chose.
The Anthropic Economic Index, which tracks how workers actually use the model, has measured directive automation — where the model performs the task and the worker reviews the output — rising from approximately twenty-seven percent of observed usage in December 2024 to approximately thirty-nine percent by early 2026. The recent quarters show modest movement back toward augmentation in some channels, but the multi-year trend is unambiguous. Automation produces redistribution. Augmentation produces lift. The aggregate zero is what you get when the dominant pattern is automation in low-experience-premium occupations and augmentation in high-experience-premium ones, mixed in the right proportions.
The World Economic Forum projects a net positive of approximately seventy-eight million jobs by 2030 across the global economy. The same projection assumes massive churn underneath the net number. The headline gain is the average of a much larger gross movement. The redistribution thesis says the movement is the story and the average is the disguise.
Winners and Losers
The winners are workers whose jobs are mostly tacit. Senior lawyers, experienced financial analysts, attending physicians, surgeons, machinists who can diagnose a vibration by listening to it. Their codified work becomes faster. Their tacit work becomes more valuable. The wage premium for experience expands because the codified portion of the job, which was the floor on entry-level wages, has collapsed.
The losers split into two groups. The first is junior workers in high-exposure professions: bootcamp graduates, recent computer science majors, paralegals, junior accountants. They cannot build tacit knowledge because the codified work that taught it has been automated. The second is workers in low-experience-premium occupations: customer service, retail support, basic data entry. Their jobs were already mostly codified, and the codified portion is now done by software.
Credential-based hiring loses status as a screening mechanism. A computer science degree certifies that the holder can perform codified software engineering. AI certifies the same thing more cheaply. The remaining value of the degree is the tacit knowledge built during four years of project work, internships, and debugging real systems — and that tacit content was always uneven across institutions.
Apprenticeship models gain status. The trades, where tacit knowledge is acquired through directly observable practice, are insulated from this pattern. So are the legacy professional structures (law firm partnership tracks, medical residency programs) that built tacit knowledge through years of supervised work on real cases. The professions that ran on this model are positioned to continue running on it. The professions that replaced apprenticeship with credentialing are exposed.
What Would Falsify This
If entry-level software developer employment in the United States recovers above its 2022 peak by the end of 2027, the bifurcation thesis fails. The recovery would mean AI created enough new junior-appropriate work to offset the codified work it displaced.
If the Dallas Fed pattern reverses in subsequent quarters — if AI exposure starts associating with wage gains in low-experience-premium occupations or wage losses in high-experience-premium ones — the codified-versus-tacit mechanism is wrong and the thesis fails.
If the Yale Budget Lab aggregate stops being zero and starts showing meaningful net job loss or gain, the redistribution story will need to absorb a new direction. The thesis can survive a downturn or an upturn. It cannot survive the aggregate splitting in a way that decomposes neatly by age and experience premium.
The aggregate has been zero for long enough that the conversation has moved on to other anxieties. The bifurcation underneath it has not moved on. Each quarter that passes without the entry-level numbers recovering is a quarter in which the apprenticeship layer remains broken, and the workers who would have built tacit knowledge in 2026 will not have it in 2036.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)