Google Trial Finds AI Improves Patent Drafting as Junior Skill Gains Split
A three-month randomized field experiment lifted assisted drafting quality, but junior lawyers showed no average gain when later tested without AI.
Edited by Tyronne Panaino
A three-month randomized field experiment reported by Google Research on October 7 found two different effects from routine AI use by patent attorneys: assisted patent drafts scored better, while later tests without AI showed gains concentrated among senior lawyers rather than a uniform improvement across the workforce. The distinction matters to law firms and other expertise-heavy employers deciding whether faster output also builds the judgment junior staff need over time.
The Google Research account covers 133 lawyers at 11 intellectual-property firms that had regular, non-exclusive business relationships with Google. Two-thirds of the lawyers at each firm received early access to a then-unreleased AI patent-writing assistant, while the remaining lawyers received training but no tool access during the study period. That design supports a useful comparison, although the sample, setting and vendor relationship narrow how far the result can travel.
Assisted drafting improved across the trial
Researchers evaluated simulated patent-drafting tasks after 10 days and again after 90 days. Independent patent professionals graded enforceability, accuracy, strategic ambiguity, completeness and clarity. The treatment group scored 0.34 standard deviations above controls after 10 days and 0.38 standard deviations higher after 90 days.
The reported improvement was mainly a shift away from poor work toward lower-middle and middle-quality scores, not an increase in exceptional submissions. Junior lawyers also completed the early drafting task faster: the source reports an 18-minute saving against a 124-minute control-group average. Those results describe output produced with access to AI; they do not by themselves establish that the lawyers had learned to perform better without it.
The unassisted test separated output from learning
After 90 days, participants completed a redlining exercise without AI. This task required them to identify and correct substantive and stylistic defects in a hypothetical patent, providing a separate measure of judgment after the tool was removed.
Across all lawyers, the treatment group scored 0.32 standard deviations above controls. Senior lawyers with at least seven years of experience drove that effect, recording a 0.45-standard-deviation advantage. Junior lawyers showed no discernible average improvement. Their scores instead spread in both directions, with more low and good outcomes rather than a consistent shift upward.
This does not prove that AI harms junior development. It shows that better AI-assisted work and durable unassisted skill are different outcomes, and that this trial did not find one uniform junior-learning effect. The study also lasted only three months, used one profession, involved a limited sample and tested a tool that was unreleased during the experiment.
What employers can take from the evidence
A workplace evaluation should measure at least two things: the quality and speed of work completed with AI, and performance on tasks where the system is absent or inappropriate. Senior staff may be better positioned to use generated material as a prompt for deeper review because they already possess the domain knowledge needed to challenge it. Junior staff may need deliberate practice, supervision and unassisted checkpoints so automation does not substitute for the work that develops judgment.
The next useful evidence would replicate the design in larger samples, other professions and longer time periods, with transparent analysis of who improves, who falls behind and which training practices change those outcomes.
Status
Learning. The randomized field experiment supports a difference between assisted output and unassisted judgment, but the evidence comes from a Google research account of a limited, Google-connected professional sample and was not independently reproduced in this run.
Sources
Update note: Last reviewed 2026-10-12. We will revise this post if the full study record, replications or longer-run results materially change the conclusion.
Sources
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.