theodorewright11/ai-workforce-exposure-datasets-public

AI Workforce Exposure — Final Datasets

Task-level AI exposure datasets covering the U.S. workforce. The collection has 19 per-version datasets and 69 cumulative datasets (88 files total).

Created for the paper "Mapping AI Exposure Across the U.S. Workforce: Evidence from Millions of AI Conversations" (Wright, Schwarze, & Boyd, 2026).

📦 Code, pipeline, and full documentation

This repo hosts only the final output files. The construction pipeline, its inputs, the methodology, and the complete column dictionaries live in the GitHub repository:

https://github.com/theodorewright11/ai-workforce-exposure-dataset-construction-public

Two task spaces for the AEI-only buckets

The three AEI-only cumulative buckets ship in both O*NET vintages, suffixed _eco2015 and _eco2025. _eco2015 preserves the native v20.1 / 2010-SOC structure and ends at 2026-02-12. _eco2025 maps the same sources onto the v30.1 / 2019-SOC backbone, which admits the June 2026 AEI release (v6.1/v6.2) and runs to 2026-05-31. These replace the earlier aei_*_usage_2025_<date> files, where the task-set marker was indistinguishable from a snapshot date.

Task time estimates

Every dataset carries two columns describing how working time is spent: time_per_day, the hours per day spent on a task, and time_per_instance, the hours a single execution of it takes. Both come from Estimating Time Spent on Work.

That study scores O*NET's task ratings file, while these datasets are built from the task statements file, which additionally contains tasks O*NET lists but has never surveyed. About 7% of (task, occupation) pairs therefore have no direct estimate and are imputed with the same fallback chain used for the O*NET frequency, importance, and relevance ratings.

Daily times are shares of a fixed seven-hour workday. Because the task lists here are longer than the ones that study scored, each occupation's daily times are rescaled to sum back to seven hours on the economy-wide baselines (final_eco_2025.csv, final_eco_2015.csv). On the individual AI usage datasets, which cover only some tasks, they sum to less than seven and give the number of hours of an occupation's working day that dataset covers. time_per_instance is the length of one task rather than a share of the day, so it is imputed but never rescaled.

MCP rating counts

The four MCP datasets (final_mcp_v1.csv–final_mcp_v4.csv) carry five additional columns, n_rating_1 through n_rating_5, giving the number of classifier ratings a task received at each level.

The MCP pipeline discards every level-1 rating and rescales the surviving levels back onto 1–5, so n_rating_2 through n_rating_5 account for all of the ratings behind that dataset's scores, while n_rating_1 records the discarded volume — roughly 40% of raw ratings, and not otherwise visible anywhere in the data.

One caveat worth carrying into any analysis: auto_aug_mean on these datasets is the mean of a task's top five ratings rather than of all of them, so it describes peak automatability rather than the centre of the distribution, and runs roughly half a point higher than a full mean would. These counts make it possible to recompute it on whatever basis a given analysis calls for.

Note on AEI v6

The June 2026 AEI release rounds usage shares to two decimals, so roughly half of listed tasks arrive at 0.00. These are censored, not unobserved: AEI suppresses tasks below 15 observations per million, so a listed 0.00 denotes a true share in [0.0015, 0.005). They are filled with 0.0027 — the mean of that interval measured from the prior release, which reports it uncensored — before shares are renormalized to sum to 100.