Preliminary research · Updated August 2026
The Human Work
Behind AI
AI systems learn from people. We follow the emerging market where labs hire doctors, lawyers, engineers, and other experts to turn human knowledge into training data.
Alberto Cavallo · Zoë CullenDraft · August 21, 2026
01 / The market
A new labor market,
visible day by day.
Daily observations from a leading platform show how demand for expertise evolves as AI capabilities are built.
Public job-board observations from November 11, 2025 through August 5, 2026. Compensation is estimated using posted rates and reported or assumed durations.
02 / Hiring waves
Demand arrives
in succession.
Most postings attain the majority of their eventual hiring within four weeks. Select a domain to inspect its peak, total hiring, and persistence at the end of the sample.
Figure 1 · Reframed
When each expertise wave peaked
Select a domain for details.
03 / Compensation
High starting pay.
Then, adjustment.
The same narrow occupations earn more on the platform on average, even though the marketplace is global. Rates on continuing listings decline as the market matures.
Figure 2
Posted pay relative to U.S. wages
Posted midpoint divided by the mean OEWS wage in the same SOC-6 occupation, weighted by hires. The vertical line marks equal pay.
Figure 3
Change in continuing-listing wages
Matched-listing indices follow the same jobs over time, separating rate revisions from changes in the mix of postings.
04 / AI adoption
Hiring timing reveals
where AI is catching up.
Later expert-hiring waves occur in occupations with a larger gap between what AI could plausibly do and what it is observably doing.
+0.41
Hiring timing and the AI adoption gap
Earlier waves are associated with greater realized AI use. The result does not simply reflect how automatable the occupation is.
Hires-weighted correlations across 20 SOC major groups. AI-use measures from the Anthropic Economic Index, March 2026.05 / Interpretation
Human expertise may be most valuable while a capability is being built—and less scarce once a corpus of training data has accumulated.
Fast accumulation
Hiring rises sharply and then subsides in many fields as labs assemble specialized knowledge.
Falling marginal value
Continuing-listing wages decline 10.1% in aggregate, consistent with lower demand for another unit of similar expertise.
Not universal
Science remains persistent late in the panel, underscoring that the lifecycle differs across domains.
Preliminary research note