Skip to contents

Import Eloundou et al. (2023) GPT-Exposure Scores as a Task-Grain Measure

Source: R/imports.R
onet_import_eloundou.Rd

Reads the occupation-level GPT-exposure table released with Eloundou, Manning, Mishkin, and Rock (2023), "GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models", and broadcasts it onto the tasks of a Task Ratings style panel, returning a task-grain onet_measure() keyed on (occupation, task). Every task inherits its occupation's published exposure. This is a thin adapter: it selects the score column, standardizes the O*NET-SOC code, joins to the panel, and records provenance. It does not rescale, average, or otherwise transform the published exposure values.

Usage

onet_import_eloundou(
  panel,
  path = NULL,
  url = onet_eloundou_url,
  score = "human_rating_beta",
  key = NULL,
  sheet = NULL,
  occupation_code = "onet_soc_code",
  task_id = "task_id",
  measure_id = "eloundou_gpt_exposure",
  measure_name = "Eloundou et al. (2023) GPT exposure",
  force = FALSE,
  ...
)

Arguments

panel

A Task Ratings style panel with an occupation column (occupation_code, default "onet_soc_code") and a task column (task_id, default "task_id"). Its distinct occupation-task pairs set the grain of the returned measure.

path

Optional path to a local copy of the exposure file. A comma or tab separated file or an Excel workbook is accepted. When supplied, no download is attempted.

url

Download URL used when path is NULL. Defaults to the pinned occupation-level occ_level.csv in the authors' public repository.

score

Name of the exposure column to use as the measure score. The published occ_level.csv carries human_rating_alpha, human_rating_beta, human_rating_gamma (human-labeled exposure at the alpha, beta, and gamma definitions) and the dv_rating_* model-labeled counterparts. Defaults to "human_rating_beta".

key

Optional name of the O*NET-SOC code column in the exposure file. When NULL, common column names such as "O*NET-SOC Code" are detected automatically.

sheet

Optional worksheet name or index used when reading an Excel workbook.

occupation_code, task_id

Names of the occupation and task columns in panel.

measure_id, measure_name

Identifiers recorded on the returned measure.

force

Logical; re-download even when a cached copy exists.

...

Additional arguments passed to onet_measure(), such as universe or weight_panel.

Value

A task-grain onet_measure object (key_type = "task") keyed on task_id and scored on the selected exposure column, with the occupation code retained. It is ready for onet_task_to_occupation().

Details

The published scores are occupation-level; broadcasting them to every task of an occupation is the structurally blind aggregate construction the source paper contrasts against task-aware measures. Tasks whose occupation has no published score are dropped with a warning. Pass a single-release panel so each task id is unique.

The exposure data are distributed by OpenAI under the MIT License. onet2r never bundles or ships the file; you must supply path or download it from url. Downloads are cached under tools::R_user_dir("onet2r", "cache") in the reference section and can be cleared with onet_cache_clear(what = "reference"). Cite the source paper when you use the scores.

The three exposure definitions follow the paper: alpha counts tasks exposed by direct model access, beta adds tasks reachable with complementary software built on the model, and gamma is the broadest definition. Choosing among them is a substantive decision left to the caller; the adapter does not endorse a definition.

Examples

# Offline: broadcast a small local extract onto a tiny panel.
extract <- tempfile(fileext = ".csv")
utils::write.csv(
  data.frame(
    `O*NET-SOC Code` = c("15-1252.00", "29-1141.00"),
    Title = c("Software Developers", "Registered Nurses"),
    human_rating_beta = c(0.63, 0.14),
    check.names = FALSE
  ),
  extract,
  row.names = FALSE
)
panel <- tibble::tibble(
  onet_soc_code = rep(c("15-1252.00", "29-1141.00"), each = 2),
  task_id = c("1", "2", "3", "4")
)
measure <- onet_import_eloundou(panel, path = extract)
measure$data
#> # A tibble: 4 × 5
#>   onet_soc_code task_id human_rating_beta measure_key measure_score
#>   <chr>         <chr>   <chr>             <chr>               <dbl>
#> 1 15-1252.00    1       0.63              1                    0.63
#> 2 15-1252.00    2       0.63              2                    0.63
#> 3 29-1141.00    3       0.14              3                    0.14
#> 4 29-1141.00    4       0.14              4                    0.14

# Online: download the published occupation-level table.
if (interactive()) {
  exposure <- onet_import_eloundou(panel)
  head(exposure$data)
}