A randomized controlled trial with 133 patent attorneys at 11 law firms found that routine AI use improves work quality across the board, but its effect on learning is not uniform. After 10 and 90 days of access to an AI patent-writing assistant, lawyers with the tool scored higher on drafting tasks than controls—by 0.34 and 0.38 standard deviations respectively. However, the gains came from fewer poor scores and more good ones, not from more excellent scores, indicating AI acts as a safety net rather than a creativity booster.
The more striking result emerged when AI was removed. On a 90-day redlining task that required unassisted judgment, lawyers who had used AI outperformed controls by 0.32 SD overall—but this was driven entirely by senior lawyers (7+ years of experience), who beat controls by 0.45 SD. Junior lawyers showed no average skill gain; instead, their scores split into more strong and more weak outcomes. This suggests AI may help experienced professionals refine their judgment, while for juniors it can either accelerate learning or create dependence, depending on how they use it.
The authors note that their study is rare in combining sustained AI use over months, a credible test of unassisted skill, and blinded expert grading. They caution that AI's short-term productivity boost may not translate into durable expertise, especially for less-experienced workers. The findings align with prior research showing AI can improve immediate output but often fails to persist once the tool is removed, though this experiment adds nuance by showing seniority as a key moderator.