Headline result: exposure predicts the decline
The main specification regresses the change in log job postings (average of post-ChatGPT quarters over the four quarters preceding the release) on occupational AI exposure, with country fixed effects and standard errors clustered by country. Exposure is rescaled to 0–1, so the raw coefficient is the log-point gap between the most and least exposed occupations. The per-SD view multiplies it by the standard deviation of exposure in the estimation sample.
Estimated effect of AI exposure, by measure
Coefficient with 95% confidence interval. Hollow markers are not significant at the 5% level.
View as table
Postings change vs exposure
Each gray point is a country×occupation pair (within-country deviations); larger dots are occupation averages. The line is the regression fit with its 95% band. Hover an occupation dot to identify it.
Event study: when did the gap open?
Coefficients on exposure×quarter interactions with occupation×country and time fixed effects (quarter 0 = 2022Q4, the ChatGPT release, is the omitted reference). The post-period decline grows each quarter. In the EU data the most exposed occupations were already on a rising trend before the release, a hiring run-up that the next section controls for. The Australian data reach back to 2016 and show no comparable pre-trend over the longer window.
View as table
Alternative explanations
The post-ChatGPT window coincides with rising interest rates, the unwinding of pandemic-era remote work and a possible reversal of earlier hiring booms. Each specification adds occupation-level controls for these drivers to the long-difference model: interest rate sensitivity estimated from Reserve Bank of Australia policy surprises, the hiring run-up from 2019 into the pre-ChatGPT baseline, and the teleworkable share of Dingel and Neiman (2020). On the EU data the run-up and teleworkability attenuate some of the estimates, most for the Demirev and Anthropic measures, and all five remain significant with the three controls entered jointly. On the Australian data the controls leave the two replicating estimates unchanged or larger.
View as table
Coefficient on each exposure measure with 95% confidence interval, one regression per cell, country (state) fixed effects, SEs clustered by country (state). Samples differ across rows because each control covers a subset of cells. Hollow markers are not significant at the 5% level.
Occupation explorer
Pick up to five occupations to compare their posting dynamics against the all-occupation average (postings indexed to 100 = average of the four quarters before the ChatGPT release) and see where each sits on all five AI-exposure measures.
Where the selected occupations sit on each exposure measure
Gray ticks show the full distribution of occupations; colored dots are your selection (0 = least exposed, 1 = most exposed in the sample).
Automation vs augmentation
Two of the exposure measures decompose into an automation component (AI performs the task) and an augmentation component (AI assists a human doing it). Entered jointly, the usage-based Anthropic measure attributes the contraction to automation on both datasets, while the product-based Demirev measure does not deliver a stable split.
Occupations by automation and augmentation exposure
Each dot is an occupation, colored by its post-ChatGPT change in postings. Hover to identify.
Joint regression estimates
Both components entered together in the delta specification, same fixed effects and clustering.
Biggest movers
The occupations with the largest relative changes in postings between the pre- and post-ChatGPT periods, colored by their AI-exposure percentile. Exposure concentrates among the biggest decliners, although some low-exposure occupations declined steeply too.
Who is hit: entry level and the most senior
Using EU vacancies broken down by required experience (EURES), the association between AI exposure and posting growth follows an inverted U: most negative for entry-level postings and for those requiring over ten years of experience.
Coefficients on exposure interacted with required-experience bins; country fixed effects, SEs clustered by country. Hollow markers are not significant at the 5% level.
Methods & data
Specifications
The long-difference (“delta”) model, estimated once per exposure measure:
where ȳ is the average number of postings in the pre / post period for occupation i in country (or Australian state) c, ξc are country fixed effects, and XAI is exposure scaled to 0–1. The event study:
with country×occupation and period fixed effects; 2022Q4 is the omitted reference period. The alternative explanations add occupation-level controls to the delta model:
Standard errors are clustered by country (and occupation in the event studies). Cells with fewer than 20 postings in either period are excluded.
Data
- EU: CEDEFOP Skills-OVATE online job advertisements, ISCO 3-digit occupations × 28 countries, quarterly snapshots (each covering a trailing 12-month window); EURES vacancy data for the experience breakdown.
- Australia: Jobs and Skills Australia, Internet Vacancy Index, ANZSCO 4-digit occupations × 8 states/territories, , mapped to ISCO via the ABS correspondence.
- Exposure measures: the AI-product exposure measure of Demirev (2024); Eloundou et al. (2023); Felten et al. (2018); the Anthropic Economic Index usage measure of Handa et al. (2025); and Webb (2022). All rescaled to 0–1 across occupations.
- Controls: interest rate sensitivity from the response of Australian vacancies to unanticipated Reserve Bank of Australia cash rate decisions in 2013–2016; the EU hiring run-up from Eurostat quarterly job vacancy rates by sector, weighted by CEDEFOP employment by occupation and sector; the Australian run-up from the IVI itself (2019 to the pre-period); teleworkability from Dingel and Neiman (2020).