ChatGPT Made Spec Reviewers Worse. Save It for Round 2.
A controlled crossover experiment with 34 inspectors found ChatGPT support was associated with about 8% lower specification defect-detection F1 while saving no time at all. The cohort that learned the task unaided first improved roughly 12% between periods against 7% for the cohort that started with the tool.
August 24, 2026 · 12 min read