Which Jobs are at Risk From AI? Evaluating Karpathy’s Exposure Dashboard

Summary

Andrej Karpathy’s AI Exposure Dashboard provides an occupational scoring metric using data from the Bureau of Labor Statistics (BLS) Occupational Outlook Handbook to assess the vulnerability of various professions to artificial intelligence (AI). This note evaluates the dashboard’s heuristic approach against findings from the formal labour economics literature. The dashboard’s results are found to directionally align with established task-based exposure models, indicating high exposure for cognitive, computer-mediated occupations and low exposure for physically embodied work. However, as an exploratory metric, the tool remains subject to significant economic limitations, notably the omission of demand elasticity, within-occupation task heterogeneity, and general equilibrium adjustments.

 Evaluating 342 BLS occupations that represent over 143.06 million US jobs, the metric identifies 38.1% of occupations and 34.3% of total employment as highly exposed to AI (defined as a score of 7 or above on a 10-point scale).

Introduction

Recent advancements in generative AI have prompted widespread efforts to quantify occupational exposure. Karpathy’s AI Exposure Dashboard functions as a transparent exploratory occupation-scoring exercise rather than a structural forecast of job loss. Evaluating 342 BLS occupations that represent over 143.06 million US jobs, the metric identifies 38.1% of occupations and 34.3% of total employment as highly exposed to AI (defined as a score of 7 or above on a 10-point scale). The project explicitly cautions that it is a development tool for visual exploration of BLS data rather than a formal economic publication. This note examines the underlying methodology and evaluates its utility and limitations for economic analysis.

How the AI Exposure Dashboard Generates its Scores

The data-generating process relies on automated text extraction from the BLS Occupational Outlook Handbook. Structured fields such as median pay, education requirements, and employment counts are tabulated, while raw pages are parsed into text descriptions.

An underlying Large Language Model (LLM), specifically Gemini Flash via OpenRouter, evaluates each occupation’s description against a custom rubric to generate an exposure score. The operative heuristic, which is a subjective scoring system, anchors the highest exposure scores (7+) to occupations that can be executed entirely from a computer within a home office. Conversely, scores approach zero (0–1) for predominantly physical, hands-on jobs located in unpredictable real-world environments. The model’s prompts explicitly instruct it to account for both direct automation and indirect labour-saving productivity effects.

High exposure indicates where AI may impact workflows first, but it is insufficient to determine whether the ultimate labour market outcome will be substitution, complementarity, reorganisation, or opening up of new demand for products and services.

Comparison with the Labour Economics Literature

The resulting exposure hierarchy exhibits strong directional agreement with formal economic studies on generative AI.

• High-exposure Occupations: The metric assigns top exposure scores (8 to 10) to heavily populated cognitive and clerical roles. These include medical transcriptionists (score of 10), customer service representatives (9), general office clerks (9), software developers (9), accountants (8), and market research analysts (9). This aligns closely with Eloundou et al. and Felten, Raj, and Seamans, who demonstrate that large portions of white-collar knowledge work—particularly design, writing, analysis, and coding—exhibit meaningful task exposure to generative models.

• Low-exposure Occupations: Occupations scoring at the absolute minimum (1 out of 10) are uniformly physically embodied tasks. Examples include janitors, construction labourers, roofers, and grounds maintenance workers. This dichotomy mirrors the consensus found in the Organisation for Economic Co-operation and Development (OECD) reviews and earlier automation benchmarks by Webb, which identify manual, place-bound, and face-to-face physical work as the least exposed to current AI capabilities. Review literature from 2025 similarly notes a convergence toward high-wage and cognitive occupations being more exposed.

Workforce reskilling initiatives should pivot away from broad technical retraining and instead focus on helping workers in highly exposed occupations re-bundle their tasks around human judgment, domain verification, client interaction, and AI supervision.

Limitations of the AI-based Exposure Metric

While the dashboard provides a highly transparent, intuitive first pass for understanding task reshaping in digital occupations, it is not a causal labour-demand model. Economists utilising such machine learning or prompt-based metrics must account for several structural limitations.

• Absence of Market Dynamics: The tool does not explicitly model the channels of the exposure risk. Crucially, the scores omit demand elasticity, latent demand, and regulatory barriers.

• Measurement Granularity: By averaging exposure at the occupation level, the metric obscures within-occupation tasks that may be differentially affected by AI exposure.

• Short-term impact only: Prompt-based LLM judgments often reproduce broad narratives in the short-term rather than equilibrating forces in the long-term. Furthermore, as literature reviews emphasise, high exposure indicates where AI may impact workflows first, but it is insufficient to determine whether the ultimate labour market outcome will be substitution, complementarity, reorganisation, or opening up of new demand for products and services.

Additionally, it should be noted that this analysis is based entirely on the US labour market. Since labour market structures differ across countries, a similar exposure analysis for the Indian non-farm labour market would therefore be a useful next step.

Ultimately, for both firms and policymakers, the primary utility of these exposure maps lies in guiding task redesign and workforce planning rather than forecasting deterministic technological unemployment.

Policy and Workforce Implications

The convergence between LLM-driven heuristics and formal empirical models reinforces several operational and policy conclusions. First, workforce reskilling initiatives should pivot away from broad technical retraining and instead focus on helping workers in highly exposed occupations re-bundle their tasks around human judgment, domain verification, client interaction, and AI supervision.

Second, the scaling layer and access to tooling are critical variables. Eloundou et al. demonstrate that when LLMs are embedded directly into software and daily workflows, the share of tasks completed significantly faster rises sharply. OECD research therefore warns that unequal access to these productivity-enhancing tools can widen inequality, even when the headline exposure is highly concentrated in well-compensated white-collar professions. Ultimately, for both firms and policymakers, the primary utility of these exposure maps lies in guiding task redesign and workforce planning rather than forecasting deterministic technological unemployment.