U.S. flag

An official website of the United States government

Department of Labor Logo United States Department of Labor
Dot gov

The .gov means it's official.
Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you're on a federal government site.

Https

The site is secure.
The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

Employment Projections
Bureau of Labor Statistics > Employment Projections > Publications

Artificial Intelligence (AI) exposure categories

BLS employment projections are intended to capture long-term structural changes in the labor market. Assessing potential impacts of new and emerging technologies, including artificial intelligence (AI), is an area of focus for the BLS Employment Projections (EP) program.

EP acknowledges that any potential employment impacts stemming from AI are highly uncertain. Yet, potential AI impacts on occupations are an important component of career planning. To complement the 2025–35 projections, BLS combined several external data sources, including various theoretical AI exposure and observed AI usage measures, to create four high-level AI exposure categories.

The supplemental AI exposure table provides information about how occupations compare to one another based on their theoretical and observed exposure to AI. Theoretical exposure signifies that AI technology could be used to assist or complete some of the work performed in an occupation. The current-evidence measures incorporate observed AI interactions that are mapped to occupational tasks or work activities. They do not directly observe whether workers in a particular occupation used AI on the job. Exposure does not imply job loss, productivity gains, automation probability, or wage effects. See the Limitations section for additional information.

While this product does not measure potential AI-related impacts on employment, it provides data users with additional information to assist with making informed career planning decisions.

Methodology

While the future employment effects of AI are uncertain, researchers have attempted to measure the impacts of AI on occupations in different ways. One way is to score each occupation’s exposure to AI based on how well the capabilities of the technology match the work performed by the occupation. That is, how much of the work, or tasks, performed by an occupation can be completed or assisted by various AI technologies?

BLS combined five external data sources to create a four-category classification of relative AI exposure for every detailed occupation for which projections are produced. Three of the external data sources are theoretical measures that score each occupation’s potential exposure to AI. The other two measures estimate occupational AI exposure using observed, real-world AI usage mapped to occupational tasks or work activities.

Table 1. Source data information
Authors Title Measuring Source of measurement O*NET data used Year of occupational and AI data Dimension type

Felten, Raj, and Seamans

 Occupational, industry, and geographic exposure to artificial intelligence: A novel dataset and its potential uses Whether AI capabilities are related to occupational abilities Humans: mTurk survey respondents Abilities 2020 Theoretical exposure

Eloundou, Manning,
Mishkin, and Rock

 GPTs are GPTs: Labor market impact potential of LLMs Whether LLM capabilities could decrease the time needed to complete occupational tasks

Humans: LLM experts

LLM: GPT-4

 Occupation-specific tasks 2023 Theoretical exposure

Eisfeldt, Schubert, Taska, and Zhang

 Generative AI and Firm Values Whether LLM capabilities could decrease the time needed to complete occupational tasks LLM: GPT 3.5 Turbo Occupation-specific tasks 2023 Theoretical exposure

Massenkoff and McCrory (Anthropic)

 Labor market impacts of AI: A new measure and early evidence AI observed exposure based on Claude data LLM: Claude Occupation-specific tasks 2023: occupational data and theoretical task ratings
2025: Claude usage		 Observed evidence

Tomlinson, Jaffe, Wang,
Counts, and Suri (Microsoft)

 Working with AI: Measuring the Applicability of Generative AI to Occupations AI applicability to occupational activities based on Microsoft Copilot data LLM: GPT-4o, GPT-4o-mini, and GPT-5 Intermediate work activities 2024: occupational data and most AI data
2025: physical vs. nonphysical task classification by AI incorporated into final measure		 Observed evidence

Note: At the time of compilation, this information was free to the public.

Data sources

The three theoretical sources do not all measure equivalent concepts. Felten, Raj, and Seamans ask survey respondents whether 10 specific AI applications are related to abilities required in the workplace. Eloundou et al. and Eisfeldt et al. consider whether a large language model (LLM) could reduce the time spent on occupational tasks by at least half.

Eloundou et al. publish six exposure measures—three based on human ratings and three based on LLM ratings. Within each set, the measures differ in how they treat tasks that would require software built around an LLM. One measure counts only tasks directly exposed to an LLM, and another measure counts both tasks that are directly exposed and tasks that require additional software along with an LLM. An intermediate measure gives software-enabled exposure half the weight of direct exposure. BLS calculates a percentile rank for each of the six measures and uses the median of those percentile ranks to create a single theoretical measure for Eloundou et al.

Eisfeldt et al. use a related but separately constructed approach: they apply the Eloundou et al. exposure framework to O*NET tasks using GPT 3.5 Turbo, assigning full weight to direct exposure and half weight to software-enabled exposure before aggregating the task scores to occupations.

The two current evidence sources both incorporate actual AI usage data into their measures; however, they measure different concepts. Anthropic calculates an observed exposure measure applying Claude usage from Claude.ai conversations and Anthropic Application Programming Interface (API) traffic, building on the conversation-to-task mapping approach developed by Handa et al. API traffic consists of requests sent programmatically to Claude by software rather than through the Claude.ai interface. Both forms of usage are mapped to O*NET tasks. Anthropic’s measure includes Claude usage for work-related activities only, requires each task to theoretically be made more efficient by AI (by using the task rating in Eloundou et al.), and gives higher weight to tasks with relatively higher automative uses than augmentative uses.

The second source of observed data is from Microsoft and measures observed Copilot data matched to O*NET’s Intermediate Work Activities. These work activities are then linked to tasks, to aggregate up to the occupation level. The measure incorporates the LLM's success rate and ensures that the AI model assisted with or performed at least a moderate fraction of the work activity.

Calculation

For each source, BLS mapped the available results for occupations to EP’s occupational classification system, which is based on the 2018 Standard Occupational Classification (SOC). The percentile rank of each occupation's score within each source was then calculated for all covered occupations. Using the percentile rank, or the percentage of scores that were strictly less than the exposure score for that occupation, puts the results from all five sources on the same scale of 0 to 1 (normalization) and changes the interpretation of the measures from a raw value to a relative value that compares the measure to other occupations within the same source.

Some source measures do not cover every occupation. After the available data were normalized using the percentile rank, missing values were imputed using a procedure that predicts a missing rank from the other available source data and the occupation’s two-digit SOC major group. Across 4,155 possible occupation-source combinations for the detailed occupations for which BLS publishes projections data (831 occupations times 5 sources), 3,944 were observed and 211 were imputed, affecting 75 occupations.

The five sources were combined by first calculating two dimensions of the data for every occupation that EP prepares projections for:

  1. The median percentile rank across the three theoretical sources
  2. The median percentile rank across the two observed current-evidence sources

After the theoretical and observed median percentile ranks were calculated for each occupation, BLS then used a clustering algorithm to group the occupations into four categories of relative AI exposure. The categories are:

  • Low
  • Moderate
  • High
  • Very high

“Low” relative AI exposure generally indicates that the requirements of an occupation do not match well with the current capabilities of AI models, and that LLMs have not been observed performing many of the occupation’s tasks. “Very high” relative AI exposure generally indicates that, compared to other occupations, a larger fraction of an occupation’s tasks can be completed or assisted by AI technology, and that LLMs have been observed performing some of the occupation’s tasks.

Limitations

The AI exposure categories depend on the occupation universe, the five input sources, the occupational crosswalks, the imputation procedure for handling missing occupation data, the normalization method, and the categorization choices described above. They should not be interpreted outside that scope.

Interpretation limits

  • An exposure category is not a forecast of employment growth or decline.
    • ”High” or “Very high” relative AI exposure does not necessarily mean employment will decline.
    • ”Low” relative AI exposure does not necessarily mean the occupation will be unaffected by AI or is safe from future technological change.
  • An exposure category is not a wage forecast.
  • An exposure category is not an estimate of the probability of AI adoption.
  • An exposure category is not a worker replacement estimate.
  • An exposure category is not a productivity forecast.
  • The exposure category is relative to other occupations. It does not imply an absolute “high” or “low” level of exposure.
  • The exposure categories do not distinguish between AI impacts from automation versus augmentation.

Measurement limits

  • AI technology is advancing rapidly. The capabilities of AI in this analysis are limited and specified for a certain time, which is already in the past:
    • The theoretical sources conceptualize AI capabilities as those available no later than mid-2023.
    • The observed exposure sources may lean towards early adopters of specific AI models.
    • Most sources are explicitly focused on language modeling and do not include image or video generation.
  • The results of classifications by LLMs are subject to the specific LLM and the specific prompt being used.
  • All five data sources treat occupations as made up of a bundle of independent occupational characteristics (i.e., tasks, abilities, or work activities), and the sources do not consider bottlenecks, complementarities, or other non-independent features of occupations.
  • Source coverage, occupational crosswalks, and missing-value imputation can affect the assignment of the AI exposure category by detailed occupation.

Data file for download

Future research and analyses

BLS continues to assess the impact of various factors, including AI, on future employment. Projections are published on an annual basis and incorporate the latest available data and analysis. The 2026–36 employment projections will be released in 2027.

Additional resources

Last modified date: August 27, 2026