AI transforms jobs through task shifts, not elimination
Whilst much debate centres on job displacement from artificial intelligence, emerging analysis suggests automation primarily transforms what workers do within existing roles, creating varied earnings effects even without widespread job losses.
A study for the Federal Reserve Bank of Minneapolis by researchers from Boston College and Arizona State University examines how large language models affect wages by changing task requirements rather than eliminating positions.
The study uses worker employment data and analyses occupational task content to identify which skills are most vulnerable.
Within highly exposed occupations such as office and administrative roles, workers specialised in information-processing tasks experience wage losses as their skills become less valuable. Meanwhile, colleagues specialised in customer-facing and coordination tasks see wage gains as work shifts toward their strengths.
Estimated average skill levels and variation across workers for different workplace tasks. Tasks at the top (like detailed manual work and meal preparation) show relatively low variation in worker skills, whilst tasks toward the bottom (like analysing systems and producing technical documentation) show much wider gaps between workers' capabilities—suggesting AI automation of these latter tasks will create more unequal outcomes depending on individual workers' strengths.
Findings show LLM-driven automation creates larger workplace changes than historical industrial robot automation. Average wages in exposed occupations rise primarily because the worker mix changes, not because existing workers gain. Three groups emerge:
- Workers who leave their transformed roles and suffer losses;
- Those who stay and benefit as tasks shift toward their skills, and;
- New workers who enter these occupations after automation removes previous barriers.
The research challenges the assumption that automation exposure means uniform wage losses. Workers in the same occupation may gain or lose depending on their individual skills, suggesting exposure measures indicate potential for change rather than guaranteed outcomes.