Google researchers David Autor and Tanya Rodchenko report that AI assistance improved patent drafting quality in a three-month randomized trial, but its effect on lawyers’ unassisted judgment differed by seniority. The experiment gave 133 lawyers at 11 intellectual property law firms access to a then-unreleased Google Labs patent-writing assistant; two-thirds at each firm received the tool, while the control group received AI training but could not use it during the study. The assistant is now part of Gemini Notebook.

After 10 days, access to AI raised drafting scores by 0.34 standard deviations, equivalent to a 10-point rise in percentile ranking against control scores. After 90 days, the increase was 0.38 standard deviations, or 11 percentile points. Independent legal experts scored five quality dimensions: enforceability, accuracy, strategic ambiguity, completeness and clarity. AI users received fewer poor scores and more good scores, with no noticeable change in excellent scores. Junior lawyers also completed the 10-day drafting task 18 minutes faster than the control group’s average of 124 minutes.

At 90 days, lawyers completed a redlining task without AI, correcting substantive and stylistic errors in a hypothetical patent. Across all lawyers, AI-access participants scored 0.32 standard deviations higher than controls, equivalent to a 9-point percentile increase. The difference came from senior lawyers with 7+ years of experience, who scored 0.45 standard deviations higher, or 13 percentile points. Junior lawyers under 7 years showed no discernible average improvement: their scores spread toward both very low and good results, with fewer mediocre scores and no increase in excellent ones.

The researchers found that junior lawyers often edited from top to bottom, spending time on introductory copy before reaching patent claims, and made surface-level word changes instead of improving the claims’ commercial scope. Some identified serious flaws but commented on them without correcting them. This pattern appeared in both groups, suggesting that three months of AI-assisted rewriting had not addressed the existing weakness. Senior AI users spent more time redlining, rebuilt claims, removed wording that might narrow legal rights and linked edits to legal doctrines. In interviews, they described using AI as a “logic auditor” that helped them question existing wording and explain structural edits.

The authors say better immediate work and lasting professional judgment are different goals, and may be in tension for junior workers. They caution that their sample was limited to patent lawyers and the trial lasted only three months, far less than the years needed to develop durable expertise. Changes in AI capabilities, lawyers’ familiarity with AI and legal-sector product adoption could also affect results if the trial were repeated. They conclude that assessing AI’s effect on professional skill requires testing performance both with tools and without them.

Google researchers David Autor and Tanya Rodchenko report that AI assistance improved patent drafting quality in a three-month randomized trial, but its effect on lawyers’ unassisted judgment differed by seniority. The experiment gave 133 lawyers at 11 intellectual property law firms access to a then-unreleased Google Labs patent-writing assistant; two-thirds at each firm received the tool, while the control group received AI training but could not use it during the study. The assistant is now part of Gemini Notebook.

After 10 days, access to AI raised drafting scores by 0.34 standard deviations, equivalent to a 10-point rise in percentile ranking against control scores. After 90 days, the increase was 0.38 standard deviations, or 11 percentile points. Independent legal experts scored five quality dimensions: enforceability, accuracy, strategic ambiguity, completeness and clarity. AI users received fewer poor scores and more good scores, with no noticeable change in excellent scores. Junior lawyers also completed the 10-day drafting task 18 minutes faster than the control group’s average of 124 minutes.

At 90 days, lawyers completed a redlining task without AI, correcting substantive and stylistic errors in a hypothetical patent. Across all lawyers, AI-access participants scored 0.32 standard deviations higher than controls, equivalent to a 9-point percentile increase. The difference came from senior lawyers with 7+ years of experience, who scored 0.45 standard deviations higher, or 13 percentile points. Junior lawyers under 7 years showed no discernible average improvement: their scores spread toward both very low and good results, with fewer mediocre scores and no increase in excellent ones.

The researchers found that junior lawyers often edited from top to bottom, spending time on introductory copy before reaching patent claims, and made surface-level word changes instead of improving the claims’ commercial scope. Some identified serious flaws but commented on them without correcting them. This pattern appeared in both groups, suggesting that three months of AI-assisted rewriting had not addressed the existing weakness. Senior AI users spent more time redlining, rebuilt claims, removed wording that might narrow legal rights and linked edits to legal doctrines. In interviews, they described using AI as a “logic auditor” that helped them question existing wording and explain structural edits.

The authors say better immediate work and lasting professional judgment are different goals, and may be in tension for junior workers. They caution that their sample was limited to patent lawyers and the trial lasted only three months, far less than the years needed to develop durable expertise. Changes in AI capabilities, lawyers’ familiarity with AI and legal-sector product adoption could also affect results if the trial were repeated. They conclude that assessing AI’s effect on professional skill requires testing performance both with tools and without them.