AI coding assistants shift engineers from creation to supervision, study finds

A new longitudinal study from researchers Annie Vella and Kelly Blincoe—published on arXiv (2605.23135)—tracked 95 professional software engineers over six months to understand how AI coding assistants (like GitHub Copilot, Cursor, and others) are reshaping work. The mixed-methods investigation surveyed 158 participants at the first time point and 101 at the second, with a matched cohort of 95. The results reveal a profound shift: 82% of engineers reported spending less time writing code, moving instead toward verification activities such as reviewing, testing, and debugging AI-generated output. The authors coin this new category of work 'supervisory engineering work'—directing, evaluating, and correcting AI contributions.

The study also uncovers a striking 'productivity-experience paradox.' While 84% of participants perceived productivity improvements at both time points, the proportion reporting a worsened developer experience in at least one dimension nearly doubled from 14% to 27%. Specifically, flow state and cognitive load deteriorated even as feedback loops improved. Engineers felt less autonomous and more burdened by oversight tasks. The findings suggest that AI coding assistants are not just tools—they are fundamentally changing the nature of software engineering, pushing developers from creators to supervisors. For organizations, this means productivity gains may come at a cost: potential burnout and reduced job satisfaction unless workflows are redesigned to support the new role.