LLMs Automate Repetitive Tasks in 80% of Knowledge Work per Exposure Studies

The Keyvan blog post argues that current LLMs target repetitive knowledge work rather than AGI-level cognition, rendering much employment unnecessary. It rejects job preservation as an end in itself and notes that top-down structures dominate daily hours. This view aligns with observed deployment patterns where firms substitute models for routine output.

Eloundou et al. (2023) measured 19% of US jobs with over 50% LLM exposure and 80% with some exposure using O*NET task mappings. GPT-4 technical reports confirm 86.4% MMLU performance on knowledge benchmarks that track common analyst and writing duties. BLS occupational data shows administrative and legal support roles declining 4-7% annually since 2023 in sectors with high model adoption.

These capabilities disrupt labor markets by targeting the grunt work that fills most salaried positions, not exceptional human traits. Historical automation followed identical paths in manufacturing, where output rose while headcount fell. Current LLM iteration cycles compress this timeline from decades to quarters.

Firms will next freeze hiring in exposed categories while expanding prompt engineering and verification roles. Output per worker metrics will rise first in software, legal review, and market research before broader reallocation.