Occupation Risk Explorer

AI exposure, workforce projections, and retirement-buffered squeeze risk for 996 U.S. occupations. Data sources: O*NET, BLS OEWS, BLS Employment Projections, Eloundou et al. (2023), Felten et al. AIOE.

Methods

This tool combines official U.S. government workforce data with two independent academic AI-exposure methodologies to assess which occupations face the most realistic risk from current AI systems. The methodology below walks through how each score is calculated, what it means, and where the data comes from.

What this tool answers

Not all AI-exposed occupations will see displacement. An occupation where 15% of workers retire or switch jobs every year can absorb a lot of AI pressure through natural attrition — firms just stop backfilling positions. An occupation where only 4% of workers leave annually can't. The tool answers a specific question: which occupations combine high AI exposure with low turnover capacity, such that AI pressure is likely to translate into actual workforce stress?

The core unit of analysis is the Standard Occupational Classification (SOC) 6-digit occupation, matched to O*NET-SOC codes. The tool covers 996 detailed occupations representing roughly 222 million U.S. workers — essentially the entire non-military civilian workforce.

AI exposure scores

Each occupation has up to six exposure measures drawn from two independent research programs. The scores complement each other: agreement across methodologies strengthens the signal, disagreement flags genuine uncertainty.

Eloundou et al. (2023)

From "GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models" by OpenAI researchers. Rates each task in an occupation based on whether GPT-4 (or an AI-augmented worker) can complete it, with three thresholds:

Eloundou α (strict)
Range: 0 to 1
Share of tasks GPT-4 can complete directly, without any tools. This is the most conservative reading — "the model alone does the work."
Eloundou β (primary)
Range: 0 to 1
Share of tasks GPT-4 can complete with modest tool access (e.g., browsing, code execution). This is the primary score used for sorting, filtering, and the squeeze calculation — it represents the realistic capability of current AI systems with standard scaffolding.
Eloundou γ (broad)
Range: 0 to 1
Share of tasks GPT-4 can accelerate under the broadest interpretation, including augmenting a human worker. This is the upper bound.

The gap between α and γ is informative: a wide gap means the occupation is more likely to be augmented (humans + AI) than automated (AI replacing humans).

Felten AIOE (Felten, Raj, and Seamans)

From a line of research mapping AI capabilities to occupational work activities. Uses a different construction: experts rate AI system capabilities, those capability ratings are matched to occupational work activities from O*NET, and the weighted result is reported as a z-score.

Felten AIOE
Range: −2 to +2 (z-score)
General AI Occupational Exposure. Positive values indicate above-average exposure; negative values below-average. Covers all AI capabilities, not just LLMs.
Felten LM
Range: −2 to +2 (z-score)
Language-model-specific exposure — isolates vulnerability to text-based AI like ChatGPT and Claude. The most policy-relevant score for the current LLM wave.
Felten Image Gen
Range: −2 to +2 (z-score)
Image-generation-specific exposure. Relevant for creative, design, and visual occupations.

Composite exposure

The composite exposure score shown in the chip and the squeeze calculation uses Eloundou β when available (covers ~922 occupations). For occupations without an Eloundou score, the composite falls back to the Felten AIOE z-score normalized to a 0–1 scale using the transformation max(0, min(1, (aioe + 2) / 4)). This ensures the composite is defined on a consistent scale even when one methodology has gaps.

Exposure tiers

The composite exposure score is binned into four tiers for the filter and visual chips:

  • High (≥ 0.5): Most tasks are AI-accessible. Major productivity or substitution effect plausible.
  • Moderate (0.3–0.5): Meaningful portions of the work are AI-accessible; partial automation or substantial augmentation likely.
  • Low (0.15–0.3): Limited AI applicability; some specific tasks could be assisted.
  • Minimal (< 0.15): Work is primarily physical, interpersonal, or requires capabilities current AI lacks.

Squeeze risk calculation

Exposure alone doesn't predict workforce stress. An occupation with 15% annual turnover (through retirements plus transfers to other occupations) can absorb significant AI pressure without layoffs — firms just slow hiring. An occupation with 4% turnover cannot. The squeeze score explicitly accounts for this.

squeeze_score = exposure × (1 − absorption) × log₁₀(employment)

Where:

  • exposure = composite Eloundou β score (0 to 1)
  • absorption = min(1, turnover / 15) — the fraction of AI pressure that annual turnover can absorb, capped at 100% when turnover is 15% or higher
  • turnover = BLS exit rate + transfer rate (percentage of workers leaving annually)
  • log₁₀(employment) = log-scale weighting so tiny occupations don't dominate, but very large occupations get some amplification

The formula rewards occupations where all three factors align: high exposure, low turnover capacity, and large workforce. The log employment term is deliberately mild — it prevents 2,000-worker occupations from appearing on par with 2-million-worker ones without requiring workforce size to dominate the ranking.

Squeeze tiers

The qualitative squeeze tier (shown in the occupation detail view) uses a slightly different rubric based on whether turnover can plausibly absorb AI pressure:

  • High squeeze: exposure ≥ 0.5 AND turnover < 8%. Retirement cannot absorb the pressure.
  • Moderate squeeze: exposure ≥ 0.5 AND turnover 8–12%. Partial absorption capacity.
  • Exposure absorbed: exposure ≥ 0.5 AND turnover ≥ 12%. AI pressure should translate to slower replacement hiring, not layoffs.
  • Lower tiers follow the exposure tiers above.

Data sources

O*NET 30.2 Database
U.S. Department of Labor · Released 2024

Provides occupation definitions, titles, and descriptions. O*NET-SOC codes map to Bureau of Labor Statistics SOC codes via the official crosswalk.

BLS Occupational Employment and Wage Statistics (OEWS)
Bureau of Labor Statistics · May 2024 estimates

Current employment counts and wage distributions (median, 10th percentile, 90th percentile) for every detailed occupation.

BLS Employment Projections 2024–2034
Bureau of Labor Statistics · September 2025 release

10-year employment projections, annual job openings, exit rates, transfer rates, and education requirements. Exit and transfer rates are the basis of the turnover absorption calculation in squeeze risk.

Eloundou et al. (2023) — "GPTs are GPTs"
OpenAI researchers · March 2023 (v1), updated August 2024

Task-level exposure ratings based on GPT-4 capability assessments. Covers 922 of the 996 occupations in this dataset. Public data repository: github.com/openai/GPTs-are-GPTs.

Felten, Raj, and Seamans — AIOE, LM, IG
Brookings / Princeton · Multiple releases through 2024

General AI Occupational Exposure plus language-model-specific and image-generation-specific measures. Covers 786 occupations, constructed from expert ratings of AI capabilities matched to O*NET work activities.

Limitations and caveats

  • Exposure is not destiny. A high Eloundou β score means "GPT-4 could theoretically do many of these tasks." It does not mean those tasks are being automated in practice right now. Real-world AI adoption depends on cost, integration effort, regulatory constraints, and firm-level decisions that exposure scores don't capture.
  • These scores don't measure robotics or physical automation. Eloundou and Felten are specifically designed to measure exposure to generative AI and language models. They do not capture industrial robotics, autonomous vehicles, self-service kiosks, precision agriculture, or automated material handling. As a result, occupations like conveyor operators, cashiers, assembly workers, and truck drivers may show "minimal AI exposure" even when they face substantial automation pressure from other technology. The tool adds a separate "Physical automation risk" assessment for occupations in production, transportation, agriculture, construction, food prep, and sales — but this is rule-based inference from BLS projection data, not a rigorous independent methodology like Eloundou or Felten. A future upgrade would integrate published robotics exposure data (Frey-Osborne, Acemoglu-Restrepo).
  • Healthcare and physical occupations need extra caution. For physicians, nurses, skilled trades, construction, transportation, and other hands-on work, aggregate exposure scores typically pick up automatable documentation and information-lookup tasks — not the physical, interpersonal, or regulated core of the job. The tool dampens squeeze-risk scoring for these occupations and flags healthcare practitioners with a specific caveat, but readers should still interpret high scores in these groups as pointing to workflow augmentation rather than workforce substitution.
  • Single-methodology occupations have weaker signal. When only Eloundou or only Felten has data for an occupation, the tool flags "Moderate confidence" because there's no cross-check between independent approaches. Agreement across methodologies is the strongest evidence; single-source scores should be treated as tentative.
  • Aggregate-level analysis only. The tool shows occupation averages. Within any occupation, there's variation — senior lawyers and junior paralegals face very different AI pressure even though both sit inside the legal occupation group. This tool can't see that variation.
  • Point-in-time data. Exposure scores reflect AI capabilities at the time of publication (GPT-4 era). AI systems have advanced since, and some exposure ratings may already be understated for the current frontier. BLS projections don't yet incorporate AI assumptions.
  • Turnover is a coarse absorption proxy. The 15% turnover threshold is a simplification. Real absorption depends on age structure, career path alternatives, and hiring pipeline dynamics that aren't captured in aggregate turnover rates.
  • The squeeze score is relative, not predictive. A high squeeze score doesn't mean "this occupation will lose X% of workers." It means "this occupation is structurally more vulnerable than most." The specific magnitude and timing of any displacement are not forecasts.
  • Not every relevant occupation is covered. New and emerging occupations (prompt engineers, AI safety researchers, etc.) may be missing from BLS occupation taxonomies. The tool inherits BLS coverage gaps.
  • Methodology disagreement is real. Where Eloundou and Felten differ, the underlying question is genuinely uncertain — different approaches to measuring AI exposure produce different rankings for specific occupations. The tool shows both rather than forcing a single answer.

What this tool is for, and what it isn't

It's for: getting a structured, data-grounded view of which U.S. occupations face the most realistic AI-driven workforce stress, and understanding why any given occupation scores the way it does.

It isn't: a career forecast, a prediction of individual job loss, a replacement for sector-specific analysis, or a policy prescription. The tool quantifies structural risk; what happens in any particular labor market depends on adoption speed, firm decisions, macroeconomic conditions, and policy choices the tool doesn't model.

Squeeze risk score = exposure × (1 − turnover absorption) × log(employment). This weights occupations where AI exposure is high, retirement/transfer can't absorb the pressure, and the workforce is large enough for the effect to matter economically. Exposure scores: Eloundou dv_beta (GPT-4 rated, primary) · Felten AIOE (general), LM (language models), IG (image generation). This tool is a what-if framework, not a forecast of individual job loss.