An NBER working paper by economists Anders Humlum and Emilie Vestergaard found that large language model adoption in the workplace is linked to occupational switching and task restructuring but has produced no net changes in hours worked or earnings1. Workers exposed to LLM-capable tasks are shifting what they do — moving between occupational categories, taking on different responsibilities — without the aggregate employment destruction that dominates corporate press releases and market commentary.
The finding echoes a pattern economists have documented across previous waves of automation. When ATMs spread through American banking in the 1980s and 1990s, the number of bank tellers did not fall — it rose, because cheaper branch operations meant more branches, and tellers shifted from cash handling to customer service and sales. James Bessen of Boston University documented this dynamic extensively: automation changes the composition of work within a job faster than it eliminates the job itself. Humlum and Vestergaard's data suggests LLMs are, so far, following the same trajectory.
The paper complicates the narratives on both sides of the AI employment debate. Companies claiming AI justifies immediate, large-scale headcount reduction cannot easily square that claim with data showing no net reduction in labour hours among LLM-exposed workers. But those who argue the technology will simply create more and better jobs face a challenge too: the paper documents occupational switching, which imposes real costs on workers who must acquire new skills, navigate unfamiliar roles, and absorb the friction of transition — even when the aggregate numbers look stable.
The gap between firm-level announcements and population-level data remains unresolved. Individual companies are cutting thousands of workers and citing AI. The macroeconomic evidence — from Oxford Economics2, the Yale Budget Lab3, and now Humlum and Vestergaard — consistently fails to find the aggregate displacement those announcements imply. Either the cuts are too recent to appear in the data, or the AI justification is running well ahead of the technology's actual capacity to replace human labour.
