Google and NBER Report AI Hinders Junior Professional Skill Development

A three-month field experiment conducted jointly by Google Research and NBER found that while AI tools improved the quality of patent draft writing, they did not contribute to the improvement of judgment skills without AI for junior lawyers. This study, involving MIT economist David Autor and the Google research team, targeted 133 patent lawyers across 11 US law firms.

Experimental Design and Tools

The experiment was conducted over three months from May 2025 to February 2026. The treatment group of 90 participants used InFlow (based on Gemini 2.5 Pro), an unreleased writing tool from Google Labs. The control group of 43 participants used existing legal tools. Evaluations included patent draft writing ability on day 10 and day 90, as well as error review (Redlining) ability on day 90 with AI usage prohibited.

Key Results: Quality Improvement vs. Skill Stagnation

• Draft Quality: The AI-using group showed overall quality improvements of 0.34 SD on day 10 and 0.38 SD on day 90. Notably, junior lawyers showed a 0.60 SD increase on day 90, demonstrating a greater improvement effect than seniors.
• Judgment Without AI: In review tasks performed without AI, seniors showed a significant skill improvement of 0.45 SD, whereas juniors showed almost no change at -0.03 SD. This suggests that while AI is effective in complementing seniors' existing expertise, it does not help juniors accumulate foundational competencies.
• Work Time: Juniors saved an average of 18 minutes with AI usage, but this did not lead to skill improvement without AI.

Study Limitations and Implications

The research team explicitly noted limitations such as Google's funding and potential conflicts of interest, the short experimental period (3 months), and insufficient sample size (91 participants completed the review tasks). Additionally, there were some discrepancies between the pre-registered plan and the actual reported content. The authors analyzed that tasks requiring interpretation and verification of AI outputs improve skills, while tasks that accept outputs as-is may deprive users of learning opportunities. Therefore, they recommended that when introducing internal AI tools, companies should separately evaluate capabilities such as code review and incident analysis performed without AI, in addition to throughput in agentic usage states.