The Counterpoint: What Stanford's Employment Data Says That Goldman Did Not
Stanford's SIEPR finds AI's employment effects are small. Goldman's RFC analysis finds net-negative hiring. Both are right. The gap between them is the story.
Two of the most influential analyses of AI's labor market effects were published 72 hours apart this month. They reach what look like opposite conclusions. They are not, in fact, in disagreement. The gap between them is the most instructive part of the story.
On July 23, the Stanford Institute for Economic Policy Research (SIEPR) released a policy brief arguing that AI's effect on aggregate employment is "likely small right now." Unemployment in the most AI-exposed occupations has risen 0.77 percentage points since 2022 — marginally less than the 0.85 point rise among the least-exposed occupations, a gap the authors attribute to broad labor market softening rather than AI-driven displacement. Employment in high-exposure occupations is stable. Firms that adopted enterprise AI saw employment grow 10 percent in the two years following adoption. Job postings for software developers — an occupation at the center of the automation anxiety — are growing faster than for other occupations.
On July 26, Goldman Sachs published Research on the Composition of Labor (RFC) scores for detailed occupations and found a net-negative relationship between AI substitution potential and hiring rates Goldman Sachs Research, July 2026. Occupations with high RFC scores — where AI can automate meaningful fractions of the work — are seeing measurable hiring declines. The finding is concentrated but statistically significant.
Two data points, same question, different answers. The difference is not a contradiction. It is a disagreement about where to look.
What Each Study Measures
The Stanford brief uses the Felten-Raj-Seamans AI Exposure Index, which scores occupations by their proximity to AI capabilities across ten application areas. This is a broad, economy-wide measure. It covers everyone from radiologists to copywriters. When Stanford finds that aggregate unemployment hasn't diverged meaningfully between high- and low-exposure occupations, it is asking: has AI created a macro-scale labor market shock? The answer, on the current data, is no.
Goldman's RFC analysis uses a different instrument: O*NET occupational descriptors scored for how replaceable each work activity is by AI systems. This is a narrower, more mechanistic measure. It does not ask whether AI is reshaping the economy broadly. It asks: does AI's capacity to substitute for specific work tasks predict whether employers hire for those tasks? The answer is yes, and the effect is visible in the hiring data, with Goldman estimating a net reduction of 10,000 to 15,000 jobs per month in exposed sectors Business Insider, July 2026.
What Each Misses
The Stanford approach understates the concentrated effects. A 0.77 percentage point unemployment rise in exposed occupations hides the composition: entry-level hiring is absorbing the vast majority of the adjustment. Research from Anthropic found a 14 percent drop in hiring rates for workers aged 22-25 in AI-exposed occupations. The Brookings Institution estimates that nearly half of the gateway-to-destination mobility pathways in the U.S. labor market are highly AI-exposed. The aggregate looks stable because the effects are concentrated at the entry point — the place least visible in broad unemployment statistics.
The Goldman approach misses a compensating mechanism. RFC captures substitution — Leontief's machine replacing the worker. But it does not capture the demand-side response — what I have previously called the Volume Multiplier: when AI lowers the cost of a service, demand for that service can expand enough to offset the per-unit labor displacement. This is Jevons paradox applied to information work, and the Stanford finding that software developer job postings are growing despite high RFC scores is consistent with exactly this effect. Goldman sees the substitution. Stanford sees the demand. Neither fully models the interaction.
The Synthesis
The two studies are consistent with a single narrative: AI is not creating a uniform, economy-wide labor market shock. It is creating concentrated effects at specific points in the occupational structure — entry-level cognitive work, routine analytical tasks, the bottom rungs of professional careers — while demand expansion offsets or masks the effects everywhere else.
This is not the "jobs apocalypse" that Anthropic CEO Dario Amodei has warned about, and it is not the "nothing to see here" that raw aggregates suggest. It is a structural shift in the composition of labor demand that existing measurement tools are poorly equipped to capture. The Stanford brief and the Goldman RFC analysis are not competitors. They are two partial maps of the same territory, and the gaps between them are where the actual mechanism lives.
The policy implication is specific. If AI effects are concentrated at entry points and in specific occupational segments, then the right response is not economy-wide social insurance — though that may eventually be needed — but targeted interventions at the gateway: apprenticeship programs, income-support for displaced entry-level workers, and data infrastructure that can track occupational composition rather than just aggregate employment. The institutions that produce the data Stanford and Goldman both rely on — the BLS, the BEA — are operating under conditions that make their outputs less reliable than they were. That is a separate concern. But it compounds this one: we are trying to measure a structural shift in labor demand with instruments that were not designed for it, while those instruments are themselves under pressure.
The question is not whether Stanford or Goldman is right. The question is whether we are building the measurement tools that could tell us. On the current evidence, the answer is no.
Sources
Stanford SIEPR. What Is Really Happening to Jobs? Separating AI Hype from Reality. July 23, 2026.
Goldman Sachs Global Research. AI Impact on Jobs Briefing. July 2026.
Hennessy, Georgia. Goldman Sachs Economist Predicts AI Displacing 15 Million Jobs. Business Insider, July 2026.
Anthropic. AI Labor Market Update. March 5, 2026.
Felten, Raj, and Seamans. Occupational AI Exposure Index. Research Policy, 2021.
Brookings Institution. Occupational Pathways to the Middle Class. 2026.
NBER. Firm-Level AI Adoption and Employment Outcomes. 2025.
