noworkleft.com

Home › Sources

Method

Sources, licences and limits

Nothing on this site is our own estimate. We join public datasets, show each separately and say how reliable the join is. This page lists them all.

The datasets

DatasetWhat it saysUsed
O*NET® Database
U.S. Department of Labor, Employment and Training Administration
Occupations and their task statements (US). The skeleton everything else is joined to.Release 31.0, fetched 2026-11-06. CC BY 4.0. This site uses data from O*NET® Database by the U.S. Department of Labor, Employment and Training Administration (USDOL/ETA), licensed under CC BY 4.0. We have modified it (merged variants, matched to other sources). USDOL/ETA has not approved or endorsed this site.
Eloundou, Manning, Mishkin, Rock: GPTs are GPTsFor each O*NET task: could an LLM alone (E1) or an LLM plus extra software (E2) cut the time by at least half, or neither (E0)? Rated by human annotators and by GPT-4 under two rubrics.Fetched 2026-11-06. MIT licence. Ratings reflect early-2023 models.
Anthropic Economic Index, labour-market-impacts filesAn occupation-level “observed exposure” score and a per-task “penetration” value from Claude usage. Defined in Anthropic’s study: it combines the Eloundou ratings with real use, giving full weight to automated and half to augmentative use.Fetched 2026-11-06. Released under CC BY (the dataset card also tags MIT; we follow the stricter CC BY and credit Anthropic). Anthropic builds Claude: this is a vendor’s own usage data, covering Claude only and mostly US use.
Indeed Hiring Lab Job Postings Index and AI TrackerDaily job-postings indices by sector and share of postings mentioning AI keywords.Fetched 2026-11-13. CC BY 4.0.

How we join them

The Eloundou labels and Anthropic’s usage values are matched to the current O*NET task statements by task ID and, failing that, by normalised wording. In the current build 99.4% of the 17,938 tasks have a human/GPT-4 label and 99.2% have a usage value; the build fails if either drops below 90%. 756 occupations at six-digit SOC level are kept (those Anthropic publishes), merging O*NET’s detailed variants and removing duplicate task texts.

The “could” share on a job page is the unweighted share of its tasks rated E1 (LLM alone) or E1 or E2, by each rater; we show the range between raters instead of choosing one. Anthropic’s observed measure only counts tasks that were also rated feasible (in our data 1,651 of the 1,654 tasks with measurable use were rated as LLM-doable by GPT-4), so “observed” is a subset of “could” by construction. We do not weight by time spent, because the public files do not give task time shares. Anthropic’s occupation score is shown as published.

What these numbers are not

Privacy

The task tool runs entirely in your browser. What you type is not sent to us, stored or logged. The site uses Google Tag Manager and Google Analytics in consent mode: until you accept, no analytics cookies are stored. Your choice is kept only in your browser. There are no accounts, forms or advertising.

Who and corrections

noworkleft.com is an independent project and is not affiliated with O*NET, USDOL, OpenAI, Anthropic, or Indeed. If you find an error, tell us which page and which number, and we will check it against the source.