Generative AI and the labour market: the disconnect between task-level capability and downstream employment

Generative artificial intelligence (GenAI) has produced rapid, well-measured gains in task-level capability since the public release of ChatGPT in late 2022, yet the transmission of those gains into downstream employment and earnings remains slow, uneven and shaped by frictions that headline coverage of the technology often understates. Linked administrative records from Denmark, Finland and Norway estimate small effects precisely on incumbent earnings and hours; large multi-country firm surveys report very small employment changes; and macroeconomic projections grounded in current task content imply modest aggregate impacts. Adjustment is becoming visible only on the margin with the fewest barriers, entry-level hiring in exposed occupations. Even there, identification is contested by an organisational co-shock, the post-pandemic shift to working from home, that is closely correlated with AI exposure. This research argues that the main object of analysis is not whether aggregate employment has fallen, but the chain of intervening stages: firm adoption, depth of deployment, task reorganisation, complementary investment, and reallocation across hiring and incumbent margins. This paper proposes a research agenda that measures those stages one by one.