The Pulse
Google Study Finds AI Raises Patent Quality but Splits Skill Gains
A three-month trial of 133 patent lawyers found that Google's AI drafting assistant improved work quality, but gains in unaided judgment appeared only among senior lawyers. Junior lawyers showed no average skill improvement after the tool w
AI.info Team · Written with AI ·
“Foundational expertise may be a prerequisite for extracting durable skill from AI-assisted practice.”
David Autor et al., authors of the NBER working paper
A three-month trial of 133 patent lawyers found that access to an AI drafting assistant improved the quality of their patent work, but did not produce the same lasting gains in professional judgment for every experience level. Senior lawyers improved on a later test without AI; junior lawyers showed no average gain, and their results split toward both higher and lower scores.
Google Research published the findings on October 7, 2026, describing a randomized experiment at 11 intellectual-property law firms. The researchers tested two different outcomes: the quality of work produced with AI available, and lawyers’ ability to review a patent without AI after three months. Separating those measures, the authors argue, helps distinguish a better assisted result from a lasting increase in skill.
AI improved patent drafts at both checkpoints
The 133 participating lawyers were randomly assigned within their firms. Two-thirds received early access to a Google Labs patent-writing assistant, while the rest received AI training but could not use the tool during the study. The firms had regular, non-exclusive business relationships with Google, a detail that matters when weighing the setting and the study’s funding.
Researchers assessed drafting after 10 days and again after 90 days, using different simulated invention materials for each exercise. Independent patent professionals rated the submissions on enforceability, accuracy, strategic ambiguity, completeness and clarity. At the first checkpoint, access to AI raised drafting scores by 0.34 standard deviations; at 90 days, the difference was 0.38 standard deviations.
Google’s post equates those results to roughly 10- and 11-percentile-point gains over the control group. The increase came from fewer poor scores and more good ones, rather than a rise in exceptional submissions. Junior lawyers also finished the 10-day task 18 minutes faster than the control group’s average of 124 minutes.
The no-AI test separated senior and junior lawyers
At the end of the trial, all participants completed a redlining exercise: they had to mark up and correct an error-filled hypothetical patent without any AI. The researchers used the task to test judgment that lawyers might need when an AI assistant is unavailable. Across all participants, the group that had received AI access scored 0.32 standard deviations above the control group.
That average concealed a sharp difference by experience. Senior lawyers—defined as those with at least seven years in practice—scored 0.45 standard deviations above their controls. Junior lawyers had no discernible average improvement. Their scores instead shifted in both directions: more strong results, but also more very low results, with fewer mediocre ones.
The researchers’ account of the task offers a possible explanation. Junior lawyers often worked from the beginning of a document, spending time on lower-stakes wording instead of the central patent claims; even when they identified serious problems, they sometimes described the flaw without making the correction. Senior lawyers were more likely to rebuild claims and explain edits through legal principles. The authors suggest that experienced lawyers may have had the foundation needed to use AI output as material for sharpening their own judgment.
Three months cannot settle a career-long question
The paper does not establish that AI prevents junior lawyers from developing expertise, or that the same pattern will appear in other professions. The researchers describe the sample as limited and the three-month observation window as short compared with the years it takes to develop professional judgment. Their results also reflect the AI tools and user familiarity of the study period; Google notes that model capabilities and legal-product use have since changed.
The study’s funding and authorship also warrant attention. The NBER working-paper page says Google provided funding for the experiment’s direct costs and lists Google employment or contractor disclosures for several co-authors. Those disclosures do not negate the trial’s results, but they are relevant context for readers assessing research conducted with firms that do business with Google.
For employers, the distinction the trial measures is concrete: AI can raise the quality of a task while leaving open whether the worker becomes better at doing it independently. In this patent-law experiment, the assisted drafts improved across experience levels; after the tool was removed, the measurable average gain belonged to the senior lawyers.