A study summary published Oct. 7 described a three-month randomized experiment involving 133 practicing patent lawyers at 11 U.S. intellectual-property firms. Access to an AI drafting assistant raised scores on simulated patent-drafting tasks; on a later task completed without AI, the average advantage was concentrated among senior lawyers, while junior lawyers showed no average gain.
The experiment compared two distinct kinds of work: drafting with AI and correcting a flawed patent draft without it. Its findings concern those tasks and the participating lawyers—not employment or the future of the profession.
How the patent-law experiment worked
Two-thirds of the lawyers at each firm received early access to a custom AI assistant for patent drafting. The control group received some training in AI use but did not get access to the study tool until the three-month trial ended.
Participants drafted patents from hypothetical inventor materials after 10 days and again after 90 days, using different simulated inventions. At 90 days, they also edited a flawed hypothetical patent without AI. This kind of detailed markup, often called redlining, involves identifying and correcting problems in a draft.
Blinded expert patent attorneys graded the submissions on enforceability, accuracy, strategic ambiguity, completeness and clarity. The reported score differences are in standard deviations, a way to express the size of a difference relative to the spread of scores. They are not percentage increases.
AI-assisted drafting scores rose at both test points
Lawyers with access to the AI assistant scored 0.34 standard deviations above the control group on the simulated drafting task after 10 days (p = 0.03). After 90 days, the difference was 0.38 standard deviations (p = 0.01).
The experiment also reported a time difference for junior lawyers on the 10-day task: they finished 18 minutes faster than the control group’s 124-minute average. That comparison applies to this specific task and group, not to patent drafting in general.
Unassisted redlining gains were concentrated among senior lawyers
On the 90-day redlining task, completed without AI, the group with earlier access to the assistant scored 0.32 standard deviations above controls overall (p = 0.04). The average difference was concentrated among lawyers with at least seven years of experience: senior lawyers scored 0.45 standard deviations above controls (p = 0.02).
Junior lawyers—those with fewer than seven years of experience—showed no average gain on the task. That group result does not mean every junior lawyer’s performance moved in the same direction.
| Task and group | Timing | Result compared with control group |
| AI-assisted simulated drafting | 10 days | 0.34 standard deviations higher (p = 0.03) |
| AI-assisted simulated drafting | 90 days | 0.38 standard deviations higher (p = 0.01) |
| Redlining without AI, all lawyers | 90 days | 0.32 standard deviations higher (p = 0.04) |
| Redlining without AI, senior lawyers (7 or more years of experience) | 90 days | 0.45 standard deviations higher (p = 0.02) |
| Redlining without AI, junior lawyers (under 7 years of experience) | 90 days | No average gain reported |
The task observations also point to a difference in how participants approached revisions. Junior lawyers often spent time on introductory copy edits or identified problems without implementing fixes. Senior lawyers with AI access more often rebuilt claims and considered how wording could narrow protection. Those patterns describe the tested redlining task, not a universal measure of legal skill.
What the findings say about patent-law work
The study’s grading criteria offer a concrete view of the work it tested: a patent draft was assessed for enforceability, accuracy, strategic ambiguity, completeness and clarity. The unassisted exercise also involved prioritizing substantive issues and carrying corrections through in the text.
The different results matter. Higher scores on assisted drafting did not translate into an average redlining gain for junior lawyers in this experiment. For senior lawyers, the advantage appeared in the task that required them to review and revise a draft without AI. The study therefore speaks to performance on particular tasks—not a single, all-purpose measure of whether AI makes lawyers more or less skilled.
The experiment did not measure hiring or job displacement, client outcomes, actual patent validity or skill development over a full career. Its results cannot answer whether AI will replace patent lawyers.
The three-month window and funding disclosures
The trial lasted three months, a short period compared with the years involved in developing durable professional expertise. Its sample comprised 133 lawyers from 11 U.S. firms. The study paper reports that Google funded the experiment’s direct costs and lists Google employment and equity relationships for several coauthors, along with a paid contractor relationship with Google.
