noworkleft.com

Home › What the studies show

Explainer

What the studies show

Three sources sit under this site. Here is what each did, in one paragraph, and what it does not support.

Eloundou et al.: “GPTs are GPTs”

Rated every O*NET task for whether a language model, alone or with added software, could at least halve the time to do it. Done by human annotators and by GPT-4. The open data lets anyone compute exposure for any occupation. Limits: it dates from early 2023 models; exposure is a capability judgement, not adoption; the two rater groups disagree noticeably, which is why we show both.

ILO Working Paper 140 (2025)

Updated the ILO’s 2023 index using a large task database, scoring about 423 ISCO-08 occupations with a mean and spread. Its headline: globally, one in four workers is in an occupation with some GenAI exposure, and 3.3% of global employment is in the highest-exposure category; clerical occupations come out most exposed. Limits: it is a measure of potential exposure, not job loss, and the scores were built from a Polish occupational database mapped to ISCO-08.

Anthropic: labour market impacts (March 2026) and the Economic Index

Combined the Eloundou ratings with how Claude is really used, weighting automated use above augmentative use, and then asked whether exposed occupations show worse labour outcomes. Findings it reports: observed coverage is far below theoretical capability (about 33% against 94% in computer and maths tasks); computer programmers are highest at 75%; higher exposure is weakly linked to lower projected growth; no systematic rise in unemployment among highly exposed workers; tentative signs of slower hiring of 22 to 25 year olds. Limits (the authors’ own): Claude usage only, US only, early results, and judgement calls in how usage is weighted. Anthropic makes Claude, so treat it as a first-party data release, not an independent audit.

Where they disagree

Not so much on the ranking (clerical, software and professional work near the top; physical, outdoor and care work near the bottom) as on what it means. The theory-based indices imply large shares of work are touchable; the usage-based index says only a slice is touched so far; the employment data says little shows up yet. All three can be true at once. That is the point of keeping them apart.

Read the originals: GPTs are GPTs data · ILO WP140 · Anthropic. Summaries here are ours; check the papers before quoting.