AI Improves Patent Drafting, but Judgment Gains Favor Senior Lawyers, Working Paper Finds

AI improved patent-drafting performance in a three-month randomized trial, but the lasting gains in unaided professional judgment were concentrated among senior lawyers rather than junior participants.

The September 2026 working paper covered 133 practicing patent lawyers at 11 U.S. intellectual-property law firms. Lawyers with access to an AI assistant improved their drafting scores by 0.34 standard deviations after 10 days and 0.38 standard deviations after 90 days.

After the AI tool was removed, the lawyers completed an unaided task involving corrections to an error-filled patent. The AI-exposed group scored 0.32 standard deviations higher than the control group overall, but the result varied sharply by experience.

Senior lawyers improved by 0.45 standard deviations on the unaided redlining task. Junior lawyers showed no average improvement, although their results became more dispersed, with more very low and good scores, fewer mediocre scores and no increase in excellent outcomes.

Junior lawyers did complete the 10-day drafting task 18 minutes faster than the control group, whose average completion time was 124 minutes. The experiment therefore showed a clearer benefit for immediate output than for independent judgment among less experienced lawyers.

The AI assistant was an unreleased Google Labs tool later incorporated into Gemini Notebook. Lawyers used it to draft simulated patents before manually reviewing the flawed patent without AI. Patent professionals assessed the work for enforceability, accuracy, strategic ambiguity, completeness and clarity.

David Autor, an MIT economist, and Tanya Rodchenko wrote that “Improving performance and building durable professional judgment are different goals.” They also wrote that “Foundational expertise may be the missing link” between AI-assisted practice and durable professional judgment.

The findings do not directly measure Gen Z workers or establish how AI affects software developers, writers or other occupations. The three-month trial also cannot determine how the technology will affect expertise over the much longer period required to develop professional judgment.

Google funded the experiment’s direct costs, and several co-authors disclosed employment or contractor relationships with the company.