U.S. employment is up 2.5% since ChatGPT launched. Computer systems design is down 5%. Wages rose 7.5% nationally and 16.7% in tech. Young workers' job-finding rates dropped 14%. Every number is accurate. None of them describe the same economy.
The Dallas Federal Reserve published a labor market study in February 2026 that contains two findings which cannot both be reassuring. First: overall U.S. employment has grown approximately 2.5 percent since ChatGPT launched in late 2022. Second: employment in the top ten percent of AI-exposed sectors has declined one percent over the same period, with computer systems design — the industry building the technology — down five percent.
The aggregate is fine. The composition is not.
The Compression Artifact
Net employment figures are compression artifacts. They take two populations experiencing opposite realities and average them into a single number that describes neither. The 2.5 percent growth includes healthcare hiring, construction, and service sectors that have barely been touched by AI automation. It also includes the sectors where headcount has been falling for two years. The net is the average of a boom and a contraction, and like all averages of bimodal distributions, it sits in a valley where almost nobody lives.
Anthropic's own research team published an analysis introducing a metric called observed exposure — measuring not what AI could theoretically automate but what it is actually being used for in practice. They found that 97 percent of Claude's observed tasks fall into theoretically feasible automation categories. Computer programmers show 75 percent task coverage. Customer service representatives are among the most exposed occupations. But here is the finding that matters: there is no systematic increase in unemployment for highly exposed workers overall.
Read that again. The occupations most exposed to AI automation are not, in aggregate, experiencing higher unemployment.
The net is fine.
Who Loses the Five Percent
The Dallas Fed study identified the mechanism hiding inside the aggregate. AI automates codified knowledge — the kind that can be written in a textbook, captured in a procedure, taught in a classroom. It complements tacit knowledge — the kind that can only be acquired through years of doing the work. The result is a labor market that simultaneously punishes the inexperienced and rewards the experienced.
The wage data makes this visible. National nominal average weekly wages rose 7.5 percent since fall 2022. In computer systems design — the sector that lost five percent of its workers — wages rose 16.7 percent. The top ten percent of AI-exposed industries saw 8.5 percent wage growth. Fewer workers, each paid more. The mechanism is complement, not substitute — but only for those who already had the tacit knowledge that AI cannot replicate.
For young workers, the numbers tell a different story. Anthropic's analysis found a 14 percent drop in job-finding rates for workers aged 22 to 25 in AI-exposed occupations since ChatGPT launched. No comparable decrease exists for workers over 25. The decline comes not from layoffs but from employers simply not hiring — the positions that would have been entry-level roles are being absorbed by AI systems that can do the codified portion of the work without the training overhead.
The median experience premium across occupations — the wage gap between experienced and entry-level workers — is 40 percent. For lawyers and insurance underwriters, it exceeds 100 percent. And occupations with higher AI exposure typically show higher experience premiums. The technology that was supposed to democratize expertise is concentrating its economic rewards among those who already have it.
The Distribution Is the Story
This journal has been tracking the constituent signals for months. The Reallocation documented companies cutting payroll to fund AI infrastructure — Amazon tripling capex while eliminating thirty thousand workers. The Apprentice captured the paradox of falling headcount and rising wages in the same sector. The Inexperience Tax asked the question the aggregate obscures: if entry-level roles disappear, where do experienced workers come from in a decade? The Decoupling traced the unprecedented pattern of record revenue coexisting with record job losses. The Attrition distinguished spectacle-driven cuts from systematic infrastructure replacement.
Each entry captured a facet. The data now reveals the geometry.
The people losing work and the people gaining wages are not just different individuals — they are structurally different populations. The Anthropic analysis showed that highly exposed workers are 16 percentage points more likely to be female, 11 percentage points more likely to be white, nearly twice as likely to be Asian, earn 47 percent more than unexposed workers, and are nearly four times as likely to hold graduate degrees. AI exposure concentrates in educated, higher-earning, knowledge-economy occupations — the exact demographic that has historically been most insulated from automation.
The previous waves of automation — manufacturing, agriculture, routine clerical — displaced workers who were already economically vulnerable. This wave targets the economically privileged. The net effect may be smaller. The political and institutional disruption may be larger.
The Trajectory Question
Every snapshot is not an analysis. The question is the direction of movement.
Two years of data show a pattern: AI augments experienced workers and excludes inexperienced ones. The complement effect is real — wages rise for those who remain. The substitution effect is real — hiring slows for those trying to enter. Both are happening simultaneously, and the net employment figure captures neither.
But two years is the GPT-4 era. The capability line moves faster than measurement cycles. If the complement-substitute boundary is drawn at the line between codified and tacit knowledge, what happens when AI begins to encode tacit knowledge? The experience premium assumes tacit knowledge is durable. If it is not — if AI systems begin to accumulate something functionally equivalent to experience — the complement effect that currently protects experienced workers becomes a substitution effect too.
The Dallas Fed found no relationship between AI exposure and wage growth across 205 occupations. That is the aggregate speaking. The disaggregated data — by age, by sector, by experience level — tells a story of bifurcation that the net conceals.
The net is the wrong number. It answers a question nobody should be asking: is AI good or bad for the labor market? The answer is yes. The useful question is: for whom, at what age, in which sector, with how many years of accumulated judgment? And that question doesn't produce a number. It produces a distribution — one that the current measurement infrastructure was not built to capture, and that the current policy infrastructure is not designed to address.
The economy added jobs. The economy eliminated pathways. Both are in the net. Neither is the net.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)