An AI patent assistant can make lawyers write better without making junior lawyers grow faster into experts. On October 7, 2026, Google Research published a working paper released through the National Bureau of Economic Research (NBER): in a three-month randomized experiment tracking 133 intellectual property lawyers, AI access clearly raised drafting quality, yet independent judgment once the tool was taken away improved markedly only for senior lawyers, while juniors showed almost no gain.
How the Experiment Was Built
Participants came from 11 intellectual property law firms. Within each firm, about two-thirds of lawyers were randomly given an then-unreleased AI patent writing assistant, now part of Gemini Notebook, while the rest received the same AI training but had to wait until after the study to use the tool. On day 10 and day 90, every lawyer drafted a patent from simulated inventor materials, graded blind by independent legal experts. Day 90 added a redlining task with no AI allowed at all: revise a document seeded with substantive and stylistic errors. That task measures exactly the judgment that makes senior lawyers valuable.
The Numbers Say Two Things
First, the good news. Drafts written with AI scored clearly higher: 0.34 standard deviations above the control group on day 10, roughly a 10-point climb in percentile ranking, widening to 0.38 standard deviations by day 90. Juniors were also faster, finishing the day-10 task 18 minutes ahead of the control group's 124-minute average. But the distribution is telling: the gains came from fewer poor scores and more good ones, with no increase in excellent scores. AI raised the floor rather than the ceiling.
Take the AI Away, and the Split Appears
On the day-90 redlining task, where nobody could use AI, the treatment group still beat controls by 0.32 standard deviations overall, but that advantage came almost entirely from senior lawyers, who outperformed controls by 0.45 standard deviations. Juniors showed no discernible improvement, and their scores polarized, with more very low and more good results at the same time. Watching how each group edited, the researchers found that seniors treated AI output as a logic auditor, using it to rebuild claims and strip language that could narrow a patent's protective scope. The typical junior habits, working top to bottom on surface wording and diagnosing serious defects in comments without actually fixing them, were not corrected by three months of AI use.
The study's value is that it measured better output and stronger people separately. For managers rolling AI tools out to a team, the implied action is plain: junior training cannot drop unassisted practice and revision exercises just because an assistant exists, or a few years later the team will find that people who can independently review work have not grown at the same pace. The authors flag the limits themselves: one profession, patent law, and a three-month window, so extrapolating to other fields calls for care.
