New data released on August 27, 2026, by the US Bureau of Labor Statistics (BLS) provides a framework for understanding how different occupations are exposed to artificial intelligence (AI). This information, part of the BLS Employment Projections (EP) program, aims to assist with career planning decisions, though it does not predict job loss or gains.
The BLS developed four high-level AI exposure categories: Low, Moderate, High, and Very High. These categories are based on a combination of theoretical AI exposure, which considers where AI technology could assist or complete work, and observed AI usage, which maps real-world AI interactions to occupational tasks. The methodology combines five external data sources, three focusing on theoretical exposure and two on observed evidence of AI use.
Occupations categorized as having Low relative AI exposure generally indicate that their requirements do not align well with current AI capabilities, and large language models (LLMs) have not been widely observed performing their tasks. Conversely, occupations with Very High relative AI exposure suggest that a larger fraction of their tasks can be completed or assisted by AI technology, and LLMs have been observed performing some of these tasks.
The BLS emphasizes several limitations in interpreting these categories. They are not forecasts of employment growth, decline, or wage effects. A high or very high exposure does not automatically mean job loss, nor does a low exposure guarantee an occupation will be unaffected by future technological changes. The categories are relative to other occupations and do not imply an absolute level of exposure. Furthermore, the analysis primarily focuses on language modeling and does not extensively cover AI capabilities like image or video generation. The theoretical sources conceptualize AI capabilities as those available no later than mid-2023, and observed exposure sources may reflect early adopters of specific AI models.
The data sources used include research from Felten, Raj, and Seamans, which surveyed whether specific AI applications relate to required workplace abilities. Other sources, such as Eloundou, Manning, Mishkin, and Rock, and Eisfeldt, Schubert, Taska, and Zhang, assessed whether LLM capabilities could reduce the time needed for occupational tasks. Observed evidence sources include Anthropic’s measure, which applies Claude usage data to work-related activities, and Microsoft’s measure, which maps Copilot data to intermediate work activities.
For each source, the BLS mapped results to its occupational classification system, based on the 2018 Standard Occupational Classification (SOC). Percentile ranks were calculated for each occupation’s score, normalizing the data across sources. Missing values for some occupations were imputed using a procedure that predicts a missing rank from other available source data and the occupation’s two-digit SOC major group. Out of 4,155 possible occupation-source combinations, 3,944 were observed, and 211 were imputed, affecting 75 occupations. The median percentile ranks from theoretical and observed sources were then used in a clustering algorithm to group occupations into the four exposure categories.
