Research Summary: Execution and Evaluation: A New Occupational Measure and Long-Run Employment Gradients

A research study introduces an occupational framework distinguishing task execution from task evaluation, arguing that artificial intelligence automates output generation far more readily than correctness verification. By classifying 19,265 O*NET task statements, the author establishes occupation-level execution shares that are distinct from standard AI exposure and routine-task intensity metrics. Analysis of a harmonized panel demonstrates that white-collar occupations with high execution shares exhibit systematically lower employment growth across recent historical periods.

Why it matters

This research provides a refined mechanism for understanding how artificial intelligence disrupts labor markets by focusing on task function rather than aggregate technical feasibility. It demonstrates that workforce vulnerability is concentrated in generative execution rather than critical oversight and judgment. Consequently, enterprise and sectoral talent strategies must prioritize human evaluative competence to sustain operational quality and workforce resilience.

Key insights

• Artificial intelligence disproportionately automates task execution rather than task evaluation due to the relative ease of generating output compared to verifying its validity.
• Traditional AI exposure metrics focus on whether a system can perform a task, failing to identify whether humans primarily perform execution or evaluative functions within a role.
• The newly developed execution-share metric is reproducible across model coders and O*NET vintages, proving distinct from conventional AI capability and routine-task intensity measures.
• The metric relies on model-based classification rather than human-validated ground truth and offers modest statistical power beyond existing O*NET evaluation activity indicators.
• Panel data indicates persistent declines or lower employment growth across execution-heavy white-collar occupations across measured observation windows.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2607.20807