Importing Published Exposure Scores • onet2r

onet2r does not ship or endorse an exposure measure. When you want to work with a published one, two adapters read the authors’ occupation-level files and turn them into the same task-grain onet_measure() object that a hand-built score would produce:

onet_import_eloundou() reads the GPT exposure table from Eloundou et al. (2023), keyed on 8-digit O*NET-SOC codes.
onet_import_felten_aioe() reads the AI Occupational Exposure (AIOE) scores from Felten, Raj, and Seamans (2021), keyed on 6-digit SOC codes.

Both are thin adapters. They select a score column, standardize the occupation code, join it to the tasks of a panel you supply, and record where the file came from. They do not rescale, average, or otherwise transform the published values, and the package never bundles the files.

Stylized values. To keep this article offline, it writes small local extracts with the published column names and made-up values. None of the numbers below are the published exposure scores.

Choose the Task Panel

The panel sets the grain of the result: every distinct occupation and task pair in it receives a score. Here it comes from the relevance ratings of a pinned release in the bundled fixtures.

tasks <- onet_archive_read(
  "30.3",
  "Task Statements",
  path = archive_303,
  release_date = "2026-05-01"
)
ratings <- onet_archive_read(
  "30.3",
  "Task Ratings",
  path = archive_303,
  release_date = "2026-05-01"
)

task_panel <- distinct(ratings, onet_soc_code, task_id)

task_panel |>
  knitr::kable(align = "l")

onet_soc_code	task_id
15-1252.00	1001
15-1252.00	1002
29-1141.00	2001

Import an Occupation-Level Score

eloundou_extract <- tempfile(fileext = ".csv")
utils::write.csv(
  data.frame(
    `O*NET-SOC Code` = c("15-1252.00", "29-1141.00", "11-1011.00"),
    human_rating_beta = c(0.60, 0.20, 0.40),
    dv_rating_beta = c(0.70, 0.30, 0.50),
    check.names = FALSE
  ),
  eloundou_extract,
  row.names = FALSE
)

exposure <- onet_import_eloundou(
  task_panel,
  path = eloundou_extract,
  score = "human_rating_beta",
  as_of = "stylized extract for this article",
  release_version = "30.3"
)

exposure$data |>
  select(onet_soc_code, task_id, measure_key, measure_score) |>
  knitr::kable(digits = 3, align = "l")

onet_soc_code	task_id	measure_key	measure_score
15-1252.00	1001	1001	0.6
15-1252.00	1002	1002	0.6
29-1141.00	2001	2001	0.2

Every task inherits its occupation’s published score. Both Software Developers tasks get the same value, because the source is occupation-level. This is the structurally blind construction that task-aware measures are usually contrasted against, so treat it as a benchmark rather than a task-level estimate. The published file offers several definitions, such as human_rating_beta and the model-labeled dv_rating_beta; score selects one, and choosing it is your decision.

Check the Source Receipt

Each imported measure records a receipt with the file’s SHA-256 digest, size, retrieval time, and any as_of label.

receipt <- exposure$metadata$source_receipt

tibble::tibble(
  field = c("provenance_status", "as_of", "file_size", "actual_sha256"),
  value = c(
    receipt$provenance_status,
    receipt$as_of,
    format(receipt$file_size),
    receipt$actual_sha256
  )
) |>
  knitr::kable(align = "l")

field	value
provenance_status	recorded
as_of	stylized extract for this article
file_size	117
actual_sha256	0b1da5799e5db23a95d892f494dfc59b7675cd43fef1f970e637d3cfcb863b91

Record the digest the first time you read a file, and pass it back as expected_sha256 in later runs. A changed file is rejected before it is parsed.

recorded_digest <- receipt$actual_sha256

verified <- onet_import_eloundou(
  task_panel,
  path = eloundou_extract,
  expected_sha256 = recorded_digest,
  release_version = "30.3"
)
rejected <- tryCatch(
  onet_import_eloundou(
    task_panel,
    path = eloundou_extract,
    expected_sha256 = strrep("0", 64)
  ),
  error = conditionMessage
)

tibble::tibble(
  check = c("Recorded digest accepted", "Wrong digest rejected"),
  result = c(
    inherits(verified, "onet_measure"),
    grepl("digest mismatch", rejected, fixed = TRUE)
  )
) |>
  knitr::kable(align = "l")

check	result
Recorded digest accepted	TRUE
Wrong digest rejected	TRUE

Roll Up and Weight

The imported object is an ordinary task-grain measure, so the usual plumbing applies.

occupation_exposure <- onet_task_to_occupation(
  exposure,
  task_ratings = ratings,
  task_metadata = tasks
)

occupation_exposure |>
  select(onet_soc_code, n_tasks, total_task_weight, measure_score) |>
  knitr::kable(digits = 3, align = "l")

onet_soc_code	n_tasks	total_task_weight	measure_score
15-1252.00	1	95	0.6
29-1141.00	1	98	0.2


weights <- onet_weight_panel_oews(
  onet_oews_national(2024, path = oews_path),
  year = 2024
)
#> Dropped 2 OEWS aggregate rows; keeping "detailed" occupations.

national <- onet_measure_aggregate(occupation_exposure, weights)

national |>
  select(-coverage, -provenance) |>
  knitr::kable(digits = 3, align = "l")

measure_id	aggregate	total_employment	covered_employment	employment_coverage_share	n_occupations	n_reference_soc
eloundou_gpt_exposure	0.347	5234530	5023300	0.96	2	2

Because every task of an occupation carries the same score, the rollup returns the published occupation score for each occupation with rated core tasks. Chief Executives have a score in the extract but no tasks in the fixture panel, so their employment stays in the denominator as unmatched, and coverage is below 1.

Compare Definitions and Sources

model_exposure <- onet_import_eloundou(
  task_panel,
  path = eloundou_extract,
  score = "dv_rating_beta",
  measure_id = "eloundou_dv_beta",
  release_version = "30.3"
)

felten_extract <- tempfile(fileext = ".csv")
utils::write.csv(
  data.frame(
    `SOC Code` = c("15-1252", "29-1141", "11-1011"),
    AIOE = c(1.10, -0.30, 0.80),
    check.names = FALSE
  ),
  felten_extract,
  row.names = FALSE
)
aioe <- onet_import_felten_aioe(
  task_panel,
  path = felten_extract,
  release_version = "30.3"
)

aggregate_measure <- function(measure) {
  measure |>
    onet_task_to_occupation(task_ratings = ratings, task_metadata = tasks) |>
    onet_measure_aggregate(weights)
}

list(
  `Human-rated beta` = exposure,
  `Model-rated beta` = model_exposure,
  AIOE = aioe
) |>
  purrr::map(aggregate_measure) |>
  purrr::list_rbind(names_to = "source") |>
  select(source, measure_id, aggregate, employment_coverage_share) |>
  knitr::kable(digits = 3, align = "l")

source	measure_id	aggregate	employment_coverage_share
Human-rated beta	eloundou_gpt_exposure	0.347	0.96
Model-rated beta	eloundou_dv_beta	0.447	0.96
AIOE	felten_aioe	0.215	0.96

The AIOE adapter joins on the 6-digit SOC code, so every O*NET detail code under one SOC inherits the same score. The two sources are on different scales: the beta scores are shares of exposed tasks, while AIOE is a standardized index. Compare orderings across sources, not levels.

Occupations without a Score

Tasks whose occupation has no score in the file are dropped with a warning rather than scored as zero.

partial_panel <- tibble::tibble(
  onet_soc_code = c("15-1252.00", "51-2092.00"),
  task_id = c("1001", "9001")
)

partial <- onet_import_eloundou(partial_panel, path = eloundou_extract)
#> Warning: Dropped 1 task with no matching exposure score.
#> ℹ 1 occupation had no score in the import file.

partial$data |>
  select(onet_soc_code, task_id, measure_score) |>
  knitr::kable(digits = 3, align = "l")

onet_soc_code	task_id	measure_score
15-1252.00	1001	0.6

Reading the Published Files

With no path, each adapter downloads the authors’ file from a URL pinned to a specific commit and caches it in the reference section of the package cache. Pinning the digest and an as_of label makes a rerun fail loudly if the source ever changes.

exposure <- onet_import_eloundou(
  task_panel,
  as_of = "GPTs-are-GPTs commit 0471612",
  expected_sha256 = "<digest recorded on first download>"
)

aioe <- onet_import_felten_aioe(task_panel)

# Remove cached reference downloads.
onet_cache_clear(what = "reference")

The Eloundou et al. table is distributed under the MIT License. The AIOE workbook is provided for research use. Cite the source papers when you use either score.

References

Eloundou, Tyna, Sam Manning, Pamela Mishkin, and Daniel Rock. 2023. “GPTs Are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models.” arXiv:2303.10130. https://arxiv.org/abs/2303.10130.

Felten, Edward, Manav Raj, and Robert Seamans. 2021. “Occupational, Industry, and Geographic Exposure to Artificial Intelligence: A Novel Dataset and Its Potential Uses.” Strategic Management Journal 42 (12): 2195–2217. https://doi.org/10.1002/smj.3286.