AI Exposure vs BLS 2024–2034 Employment Growth
Headline statistic
Per massenkoff-mccrory-labor-market-impacts-2026 Figure 4 and accompanying regression:
A regression at the occupation level weighted by current employment finds slope = −6.07 (SE 1.32), R² = 0.027. For every 10-percentage-point increase in observed AI exposure, BLS projected 2024–2034 employment growth drops by 0.6 percentage points.
The relationship is weak in R² but statistically significant and survives the BLS’s own independently constructed projections.
Critically: no equivalent correlation exists when using Eloundou et al. β (theoretical capability) alone. Realized usage adds predictive value the theoretical measure lacks.
What the scatter shows
Binned scatterplot with 25 equally sized bins; small squares mark named example occupations:
| Occupation | Approx. exposure | Approx. projected 2024–34 growth |
|---|---|---|
| Software developers | ~30% | +14% (top outlier) |
| Electricians | ~0% | +10% |
| Registered nurses | ~10% | +5% |
| Lawyers | ~18% | +3% |
| Accountants | ~30% | +4% |
| (avg trend) | 50% | ~0% |
| Customer service reps | ~70% | −2.5% |
| Cashiers | ~10% | −6% |
The trend line is clearly downward sloping through 25 binned points, despite the substantial residual variance (R² = 0.027 means 97% of variance is not explained by exposure).
Interpretation
What the result supports
- Independent validation of observed-exposure-measure. BLS analysts (separate methodology, separate data, separate institution) produce growth forecasts that align directionally with Anthropic’s usage-weighted exposure. Two independent measures pointing at the same occupations is meaningful triangulation.
- Diffusion-aware exposure beats theoretical-only exposure. The same data, scored with Eloundou’s β alone, produces no significant correlation. Usage data is the missing layer. See theoretical-vs-observed-capability-gap.
- Software developers are the strategic outlier. High exposure and high projected growth. The most plausible reading: developers are the workers who use AI most productively while their labor demand grows fastest (because the work AI enables — building more software — itself demands more developers, Jevons-Paradox-style). See jevons-paradox-ai.
What the result does not support
- It is not a forecast of AI’s labor impact. BLS projections are themselves model-based and may already partially price in AI deployment. The correlation could reflect what BLS analysts expect rather than what will happen.
- The R² is low. 97% of cross-occupation growth variance is not explained by AI exposure. Other forces (demographic shifts, demand reallocation, business-cycle effects) dominate.
- The “−0.6 pp per 10 pp” magnitude is modest. Customer service reps at 70% exposure correspond to ~−4 pp adjustment; even a tripling of the slope leaves most occupations within their historical band.
Why this matters to Where Value Lands
- Bridges the empirical and projective frames. The exposure measure is now anchored to an independent labor-economist forecast, not merely Anthropic-internal usage data.
- Confirms the diffusion-vs-innovation frame at the BLS horizon. Analysts are projecting that the diffusion will deepen and affect employment, even if current unemployment data shows no signal (unemployment-did-exposed-workers).
- Updates H2_u-curve-of-value. Software developers as a high-exposure / high-growth outlier is a top-of-U story: developers occupy the L4–L6 application-layer interface and capture the surplus from automation rather than bearing the cost.
Caveats
- BLS projections have a mixed track record. The source acknowledges that government growth forecasts have “added little predictive value beyond linear extrapolation of past trends.” Massenkoff (2025) is cited.
- Endogeneity. BLS analysts read AI news; their projections may already incorporate the same exposure signal Massenkoff & McCrory measure. The independence of the two measures is partial, not clean.
Related
- massenkoff-mccrory-labor-market-impacts-2026 — source.
- observed-exposure-measure — the measure scored.
- eloundou-beta-exposure — the theoretical-only measure that does not correlate.
- most-exposed-occupations — the top-10 list driving the regression.
- theoretical-vs-observed-capability-gap — why theoretical alone fails.
- diffusion-vs-innovation — the frame that explains why usage-weighting matters.
- H2_u-curve-of-value — the strategic outlier story.
- jevons-paradox-ai — software developers as the demand-expansion case.
- capital-labor-divergence — surplus-distribution story.