Occupational AI Exposure Score (OAIES)

Updated 31 January 2026

OAIES is a metric that measures the weighted share of occupational tasks AI can potentially perform, using task-level granularity.
It employs a five-stage AI capability framework to capture rapid, nonlinear advancements in AI performance across occupations.
Empirical analyses demonstrate that higher OAIES values correlate with employment declines, differential demographic impacts, and evolving task structures.

The Occupational AI Exposure Score (OAIES) is a class of index quantifying the degree to which tasks within specific occupations can be performed by current or near-future artificial intelligence systems. The OAIES framework is foundational for empirical research on labor market impacts of AI, enabling dynamic, occupation-level measurement of “exposure” resulting from the rapid evolution of generative AI technologies. OAIES instruments have now become central in high-frequency labor-market studies, AI risk analyses, and workforce planning initiatives.

1. Formal Definition and Measurement Principles

At its most general, the OAIES measures, for each occupation, the weighted share of its constituent tasks that state-of-the-art (or projected) AI systems could in principle perform. Formally, for a given occupation $o$ and stage $S$ of AI technical capability:

$\text{OAIES}_{o,S} = \sum_{t\in T(o)} w_{o,t} \cdot e_{o,t,S}$

Here:

$T(o)$ is the set of fine-grained tasks O*NET attributes to occupation $o$ .
$w_{o,t}$ is the normalized “relevance” or “importance” weight of task $t$ within occupation $o$ , with $\sum_{t} w_{o,t}=1$ .
$e_{o,t,S} \in [0,100]$ represents the percent of task $t$ that generative AI at stage $S$ can perform, as assessed by state-of-the-art LLMs using structured prompt-engineering protocols.

This construction yields a stage- and occupation-specific exposure score, $OAIES_{o,S}$ , mapped onto a $[0,100]$ or normalized $[0,1]$ scale, which is estimated repeatedly over time as AI system capabilities advance (Dominski et al., 11 Jul 2025).

2. Five-Stage AI Capability Framework

To account for rapid, nonlinear jumps in AI technical reach, the OAIES framework segments technological progress into five qualitative stages, each anchored to public releases of generative AI:

Stage	Description	Representative Models
1. Pre-LLM	Classical ML and NLP (pre-Nov 2022)	word2vec, sklearn, pre-GPT
2. Early LLM	First public LLMs (Dec 2022 onward)	ChatGPT v1
3. Multi-modal LLM	LLMs with text + vision (Oct 2023 onward)	DALL·E 3 in ChatGPT
4. Reasoning Models	LLMs with advanced chain-of-thought and problem-solving (Dec 2024)	GPT-4o, Claude 3.5 Sonnet
5. Agentic AI	Autonomous agents, end-to-end workflow automation	Forthcoming; no fixed release

For each occupation, OAIES is computed for every stage, allowing the study of differential exposure trajectories (e.g., $\Delta \mathrm{OAIES}_{o, S3-S1}$ ) as successive model classes reach new performance frontiers (Dominski et al., 11 Jul 2025).

3. Data Sources and Task-to-Exposure Pipeline

OAIES construction relies on task-level occupational taxonomies, LLMs, and labor market microdata:

Tasks: Extracted from O*NET, with typical occupations mapped to 20–30 detailed work activities and associated relevance weights.
Scoring Protocol: For each $t$ , prompt LLMs (e.g., OpenAI GPT-4o, Anthropic Claude 3.5 Sonnet) to estimate the percent of subtasks they could perform at stage $S$ , returning $e_{o,t,S}$ and a qualitative confidence rating.
Aggregation: Compute the relevance-weighted mean across all tasks as described above.

This methodology is tractable for large occupation-task matrices ( $\sim$ 20,000 tasks, 1,000+ occupations) and has been shown to yield highly correlated results across independent LLMs (e.g., ChatGPT vs. Claude, Pearson $\rho \sim 0.94$ at Stage 3) (Dominski et al., 11 Jul 2025).

4. Dynamic Linkage to Labor Market Outcomes

A key innovation in recent OAIES work is the real-time linkage of dynamic exposure scores to labor market panel data. Using periodized (six-month) labor market microdata from the US Current Population Survey (CPS), OAIES is matched at the occupation level to outcomes such as:

Log employment
Unemployment rate
Intensive work margins (hours worked, share full-time)
Secondary job holding

The main regression specification is:

$\Delta Y_{o, P4–P2} = \alpha + \beta \Delta\mathrm{OAIES}_{o, S3–S1} + X_{o,P2}\Pi + \mathrm{TaskIndices}_o\,\Gamma + \epsilon_o$

where $X_{o,P2}$ includes demographic controls and $\mathrm{TaskIndices}_o$ encodes occupation-level task intensities. Analysis across 2.8 million person-months reveals that higher OAIES is robustly associated with reductions in employment, increased unemployment, and diminished work hours, with the effects stratified across demographic and task-content lines (Dominski et al., 11 Jul 2025).

5. Interpretive Findings and Comparative Statics

Empirical estimation using the OAIES yields several key results:

Negative Extensive and Intensive Labor Margins: A 10-point OAIES increment predicts a 5–8 percentage-point employment decline and a significant increase in unemployment (statistically significant $\beta<0$ for log employment, $\beta>0$ for unemployment).
Differential Demographic Impacts: Effects are more pronounced for older and younger workers, men, and college-educated individuals. High-education groups show less employment loss but greater change in work intensity and structure.
Task-Content Heterogeneity: Occupations heavily reliant on complex reasoning and problem-solving experience sharper declines in full-time work and overall employment, while manual physical occupations remain largely unaffected in current exposure domains.

These patterns indicate AI-driven labor shifts along both extensive and intensive employment margins as a function of OAIES (Dominski et al., 11 Jul 2025).

6. Methodological Innovations and Extensions

Several advances distinguish the OAIES approach:

Multi-stage, self-updating framework: Allows recalibration and extension to new AI capabilities as the technological boundary evolves.
Task-level granularity: Exposure is measured at the task, not occupational title, resolution, enabling more precise mappings to labor-market shifts.
Empirical Robustness: High correlation between independent LLMs and across similar exposure metrics; sensitivity analyses confirm stability of rankings.
Integration with Demographic and Task-Intensity Controls: Possible to disentangle compositional from technical exposure shifts.

OAIES thus functions as a high-frequency, transparent, and empirically validated measure for both academic study and policy monitoring of the labor impact of advancing AI (Dominski et al., 11 Jul 2025).

7. Contrasts with Other Exposure Metrics and Limitations

Relative to traditional automation indices, the OAIES exhibits several key differences:

Technical Feasibility over Adoption: OAIES quantifies technical exposure or the “potential substitution set,” rather than realized adoption or displacement.
Temporal Dynamics: Unlike static “automatability” indices, OAIES incorporates non-linear jumps as new model classes come online and can be matched to near real-time labor outcomes.
Limitations: OAIES does not address macroeconomic feedbacks (e.g., demand-induced job creation), nor does it guarantee realized automation. Self-assessment by LLMs may over- or under-estimate practical capability for edge-cases; however, robust prompt engineering and cross-model validation mitigate some of these concerns.

Further research is ongoing on integrating OAIES with adoption data, firm-level diffusion, and downstream wage and task content changes (Dominski et al., 11 Jul 2025).

References: