How were these figures calculated?

The US map covers 105 occupational groups based on the Bureau of Labor Statistics Standard Occupational Classification (SOC) 2018. Together these occupationsrepresent about half of US employment (BLS Employment Projections, 2025, including self-employed).

The 105 occupations span all 22 SOC major groups, selected by employment size and sector coverage. Together they represent the occupations most Americans work in and where AI impact is most consequential in terms of aggregate employment.

The BLS SOC 2018 contains 867 detailed occupations. This dataset covers 105 of them across all major sectors. The remaining 762 are not included. Niche specialisations, small occupational categories and occupations below the OEWS reporting threshold are excluded.

The international crosswalk maps each SOC occupation to an ISCO-08 code (International Standard Classification of Occupations), enabling comparison with the Dutch, German, UK and other country editions of this map. Crosswalk source: BLS SOC 2010/2018 to ISCO-08, supplemented with ILO correspondence tables.

The Janssen Practical Exposure (JPE) is the primary scoring methodology used across all AI Exposure Maps on this platform. It was developed by Simon Janssen, CTO at HappyNurse, to measure the practical, real-world impact of AI on occupations — as opposed to theoretical capability overlap.

JPE scores are estimated using a structured LLM evaluation process:

1. Input: For each occupation, the evaluator considers: core daily tasks, typical work environment, required skills, interaction patterns, and current AI tool adoption in the field.
2. Core question: “What percentage of this occupation’s core work can currently be augmented, automated, or fundamentally changed by AI systems available today or within 3-5 years?”
3. Scale: Integer 1-10 (whole numbers, not decimals).
4. Cross-validation: Every JPE score is cross-validated against three independent academic sources: Felten AIOE index (capability overlap), Georgieff & Hyee (2021) AI exposure indicators, and Frey & Osborne automation probability. Scores deviating more than 2 points from the academic consensus are reviewed and justified.

Felten measures “can AI do this task?” (theoretical capability). JPE measures “is AI actually changing this work?” (practical disruption). The full comparison with examples is in the JPE vs Felten AIOE section below.

JPE was developed by Simon Janssen (2026) as part of the Dutch AI Exposure Map project. The methodology is designed for cross-country comparability: the same rubric is applied to Dutch (CBS BRC 2014) and US (BLS SOC 2018) occupations.

Citation: Janssen, S.D. (2026). Janssen Practical Exposure (JPE): A practical AI exposure scoring methodology. simondjanssen.nl/en/ai-map/us/methodology#jpe

The primary AI score on the US map is JPE, the same as on the Dutch map. Alongside it, the US map shows the Felten, Raj & Seamans (2023) AI Occupational Exposure (AIOE) index as an academic reference score.

What AIOE measures: The AIOE index maps 10 dimensions of AI capability (as defined by the AI Progress Measurement project at the Electronic Frontier Foundation) to the 52 work abilities defined in O*NET. Each occupation in O*NET has an importance and level rating for each ability. The AIOE score for an occupation is calculated as the weighted overlap between what current AI can do and what the occupation requires.

The raw AIOE scores range from approximately −2.67 to +1.53. For display on this map, scores are linearly normalised to a 1–10 scale using the formula:

A score of 10 indicates the occupation relies most heavily on abilities that AI currently performs well (language, pattern recognition, information retrieval). A score of 1 indicates the occupation relies primarily on abilities where AI has little capability (physical dexterity, spatial reasoning under uncertainty, complex social navigation).

In September 2026 we rebuilt the US data on official BLS statistics. The JPE score per occupation did not change. One duplicate was removed: Mental Health Counselors (old code 21-1014) has been part of Substance Abuse, Behavioral Disorder, and Mental Health Counselors (21-1018) since SOC 2018 and was counted twice, so the map now has 105 occupations. Three SOC codes that do not exist in SOC 2018 were corrected (19-3031 to 19-3033, 11-2031 to 11-2032 and 11-2033, 29-1228 to 29-1229).

• Employment: BLS Employment Projections 2025-35 (Table 1.2), employment in 2025 including self-employed, summed over the SOC 2018 codes of an occupation and rounded to 1,000.
• Growth 2021-2025: OEWS national estimates, May 2021 to May 2025 (wage and salary jobs). There is no series before 2021, because of the switch from SOC 2010 to SOC 2018 in 2019-2020.
• Salary: OEWS May 2025, median gross annual wage; weighted by OEWS jobs where an occupation spans several codes.
• Jump warning: where the OEWS figure changed by 25% or more between May 2023 and May 2025, the page flags the growth figure and the forecast as extra uncertain. This affects display only.
• BLS AI exposure category: shown as context next to the JPE score (BLS AI exposure categories). It is not an input of the forecast.

The US forecast uses the same engine as the Dutch and UK maps (JPE-prognose v3). It starts from the national occupational projection and adds AI and robotics as separate layers. The design was fixed before the data was recalibrated and before the backtest.

• Base: BLS Employment Projections 2025-35. The 10-year projection is converted to five years: (employment 2035 / employment 2025)^1/2 − 1.
• AI layer: −0.25 percentage points per year for every JPE point above 4, over five years. Only where BLS names no AI factor for the occupation (Employment Projections, Table 1.12).
• Robotics layer: the midpoint of an expert range, only when physical automation starts scaling up in the US before 2030 (for example robotaxis, autonomous trucks, delivery robots, warehouse automation) and BLS names no automation or autonomy factor for the occupation.
• Forecast: (1 + base) × (1 + AI) × (1 + robotics) − 1, the midpoint.
• Band: base ±3 percentage points (an assumption), combined with the favourable and unfavourable AI slope (−0.1 to −0.5) and the robotics range, rounded outward to 5 percentage points. The band matters more than the midpoint.

Why the conditions. Since the 2023-33 edition, BLS builds AI effects into its projection for occupations where there is evidence for them (Monthly Labor Review, February 2025); in the 2025-35 edition that applies to 48 occupations. Adding our own AI or robotics layer on top would count the effect twice. Of the 105occupations on this map, BLS already includes AI for 11 and automation for 15. There the page says the effect is already included in the BLS projection, and no separate layer is added.

Deliberate differences from the Dutch model: no rule for structural decliners (after the SOC change there is only one consistent period), no starter signal and no labor shortage indicator. Trend and risk labels are the same as in the Netherlands and the UK.

Backtest and blind check: tested on the BLS Employment Projections 2023-33 against actual OEWS growth from May 2023 to May 2025 (105 occupations, error in percentage points per year, weighted by employment): 2.18 for the new model, 2.26 for the old model, 2.41 for "every occupation grows with the US" and 2.52 for "no change". The correlation with actual growth is 0.42 (old model 0.39). The new model has the lowest error, but the gain over the old model is small, because that model also leaned on BLS growth. A blind check of 15 random occupations against independent US sources found 9 plausible and 6 doubtful, none unlikely. The pattern: occupations with a strong, persistent trend are pulled too close to zero (physical therapists and security guards probably grow faster; cashiers and graphic designers probably shrink faster). The design was fixed before both tests: design document (Dutch).

The US and Dutch editions share the same JPE scoring rubric, but the context in which AI adoption plays out differs substantially.

The AI Exposure Map uses two scoring systems side by side. The primary score on the Dutch map is the Janssen Practical Exposure (JPE), developed for this project. The US map uses JPE as its primary score as well, and shows the academic Felten AIOE as a reference.

Why both? JPE is used for the treemap and career advice (practical decision-making). Felten is shown as an academic reference for transparency and cross-validation.

Note: JPE scores are LLM-estimated. If they are systematically biased, everything built on them is too. Cross-validation against Felten AIOE, Georgieff & Hyee, and Frey & Osborne mitigates but does not eliminate this risk.

• This dataset covers 105 of 867 SOC occupations. It represents the largest employers across all major sectors, not the full US labour market. Very small occupational categories and niche specialisations are excluded.
• The AIOE index is based on O*NET ability ratings as of 2023. It does not capture capabilities added by generative AI after that date (GPT-4 Turbo, Claude 3, Gemini 1.5 etc.). Current AI capabilities likely exceed those modelled.
• The AIOE index maps AI to abilities, not tasks. Two occupations with the same AIOE score may differ substantially in how much of their daily work is actually automatable. The score is an approximation.
• The 2030 forecast is a modelled scenario with a band, not a prediction. It leans on the BLS projection, so errors in that projection carry over. BLS already builds AI and automation effects into its projection for some occupations; where it does, no separate layer is added on top.
• OEWS covers wage and salary jobs only; the employment figure from BLS Employment Projections includes self-employed. Growth 2021-2025 and the job count therefore come from different bases.
• Salaries are OEWS May 2025 medians. Within occupational groups, the spread can be very large, especially in healthcare, legal and technology.
• Not all change in these occupations is AI-related. Cashiers face self-checkout adoption; office clerks face workflow software. These trends predate generative AI.

License: Creative Commons Attribution 4.0 International (CC BY 4.0)

Cite as: Janssen, S.D. (2026). US AI Exposure Map: 105 occupational groups analysed for AI impact. simondjanssen.nl/en/ai-map/us. Retrieved [date].

AI score: JPE (Janssen, 2026). Reference score: Felten, E., Raj, M. & Seamans, R. (2023). AIOE index. github.com/AIOE-Data/AIOE

Employment data source: Bureau of Labor Statistics, U.S. Department of Labor: Employment Projections 2025-35 and OEWS May 2021 and May 2025.