Google Research Finds AI Speeds Work but Halts Junior Judgment Growth

The 90-Day Field Experiment on Assisted Work

Published in collaboration with the National Bureau of Economic Research, researchers David Autor and Tanya Rodchenko examined 133 professionals across 11 firms over 90 days using an unreleased Google Labs tool (now part of Gemini Notebook). Assisted drafting tasks demonstrated persistent speed and consistency improvements: drafting scores rose by 0.34 standard deviations at 10 days and 0.38 standard deviations at 90 days. Junior practitioners cut drafting time by 18 minutes compared to the 124-minute control baseline.

The Unassisted Redlining Collapse in Junior Judgment

At 90 days, participants performed an unassisted redlining task without AI support, graded across enforceability, accuracy, strategic ambiguity, completeness, and clarity. While seniors demonstrated a 0.45 standard deviation gain over control seniors, juniors exhibited no average skill gain. Instead, junior distribution bifurcated: the volume of very low scores increased, mediocre scores dropped, and excellent scores remained flat.

The Diagnose-Without-Execute Trap

Qualitative analysis identified distinct junior failure modes. Juniors spent disproportionate time sequentially editing low-stakes introductory sections and swapping surface synonyms. Crucially, when detecting serious errors, juniors adopted a 'diagnose-without-execute' approach—leaving advisory comments rather than implementing structural corrections, a fundamental deficit that 90 days of continuous AI rewriting failed to remediate.