AI and the Labour Market, Signal, Noise and Structural Change

Narrative

Foundational post-ChatGPT research separates potential exposure from realised adoption. Eloundou, Manning, Mishkin and Rock’s 2023 GPTs are GPTs estimates task exposure rather than job loss, whereas Humlum and Vestergaard’s Danish linked survey-register work measures workplace adoption and administrative earnings and hours. Bick, Blandin and Deming find rapid US take-up, but only a small share of aggregate work time assisted, while Brynjolfsson, Li and Raymond’s staggered rollout establishes a causal productivity gain for customer-support workers, concentrated among novices and lower-performing staff.

The strongest evidence on labour demand is mixed and remains preliminary. Liu, Wang and Yu’s 285-million-posting difference-in-differences design reports a relative 9% decline in postings for highly substitutable occupations after November 2022, growing over time and concentrated in administration and professional services; Chandar’s CPS study finds no corresponding average employment or earnings divergence across highly exposed occupations through early 2025. Several studies instead identify adjustment in hiring composition: Hosseini Maasoum and Lichtinger use 65 million US résumés to associate firm-level GenAI integration with slower junior hiring, while Mahieu finds a roughly 23% fall in entry-level vacancies between high- and low-exposure occupations in Flanders, with no comparable experienced-worker decline.

Experimental productivity evidence supports work redesign rather than a settled conclusion about aggregate displacement. Dillon, Jaffe, Immorlica and Stanton’s six-month experiment across 66 firms found two fewer weekly email hours among users but no detected change in task quantity or composition. Ni and colleagues’ Alibaba experiment finds faster service and better customer-rated quality, especially for lower performers, but no objective-quality improvement and deterioration among top agents. METR’s HCAST and RE-Bench evidence documents rising autonomous software and AI-research task capability, but it is a capability benchmark, not evidence that employers have substituted agents for workers; METR’s own developer RCT update cautions that benchmark success can fail to produce mergeable work in real repositories.

The most credible interpretation through August 2026 is therefore one of early, uneven reorganisation. Administrative records in Denmark rule out earnings or recorded-hours effects above 2% two years after ChatGPT despite widespread adoption, whereas online vacancies and résumé data suggest that firms may first reduce or redesign junior recruitment. Job-posting analyses cannot by themselves distinguish AI causation from the post-pandemic correction, high interest rates, sectoral weakness or changing recruitment practices. Biswas’s six-country software-posting study makes that disconfirming point directly: developer-posting contraction began after the 2022 peak, roughly two years before the late-2024 rise in AI-related demand.