Without AI, only senior patent lawyers improve on average

Give a lawyer an AI drafting tool and the drafts get better. Take the tool away three months later and ask whether the lawyer got better too, and the answer depends on how much the lawyer knew going in. That is the central finding of a field experiment that Google Research described on Wednesday and that is written up in a National Bureau of Economic Research working paper dated September 2026.

The study was funded by Google, tested a Google product and lists several Google employees among its authors. Its design still answers the question more directly than anecdotes can: random assignment, a control group, outside graders and a test taken without the tool.

The researchers recruited 133 patent attorneys at 11 U.S. intellectual-property firms. About two-thirds, chosen at random, got early access to an AI patent-writing assistant from Google Labs that was then unreleased and, according to the blog, is now part of Gemini Notebook. The rest received some training in using AI but were kept off the tool until the three-month study ended. The study was pre-registered with the American Economic Association’s trial registry, according to the NBER listing.

Lawyers drafted patent material at 10 days and again at 90 days. At 90 days they also did a “redlining” task — reviewing and correcting patent documents — without any AI help. Independent legal experts graded the work on five dimensions: enforceability, accuracy, strategic ambiguity, completeness and clarity.

Every effect the researchers report is a gap between the lawyers who had the tool and those who did not, measured in standard deviations of the graded scores. With the tool in hand, the treated group’s drafts beat the control group’s by 0.34 standard deviations at 10 days — about 10 percentile points, by the blog’s conversion — and by 0.38 at 90 days. The paper’s abstract says junior attorneys gained more from the tool on that task.

The redlining test, done without the tool, is where the result splits. As a whole, the treated group beat controls by 0.32 standard deviations, about 9 percentile points. But that average was carried by experience. Lawyers with seven or more years in practice beat their controls by 0.45 standard deviations, about 13 points. Lawyers with less experience showed no average improvement.

The juniors’ scores did not simply stay flat. In the blog’s description, some improved, “hinting at AI’s ability to leapfrog learning,” while others spread toward the bottom of the distribution. The blog shows the split in histograms; neither it nor the abstract says how many junior lawyers ended up better or worse.

The blog sets the study against earlier experiments with software engineers, management consultants, high school students and clinicians, which it summarizes as finding that AI works like a “temporary exoskeleton”: it lifts output while it is worn, and the lift often does not stay once it is removed. The patent study gives a more divided answer. For experienced lawyers, the gain carried over to work without the tool; for less experienced ones, on average it did not. The abstract’s conclusion is that foundational expertise may be a precondition for turning AI-assisted practice into lasting skill.

The authors are candid about what the study cannot show. The sample is small, which they attribute to the cost of top-tier lawyers’ time. Three months is short next to the years it takes to build expertise. And AI models have improved since the trial, so a rerun today might come out differently.

Two more limits follow from the design. Patent drafting is a narrow, highly structured kind of legal writing, and the result may not carry over to work with looser standards of quality. And the “junior” group is everyone under seven years of experience — practicing attorneys, not the students or trainees at the very start of a career, where the question of learning is sharpest.

Then there is the question of who is asking. Google is the main funder, several co-authors disclosed full-time employment and equity at Google, the tool tested is a Google Labs product, Google Labs staff partnered on the experiment, and the blog says the 11 participating firms do regular, non-exclusive business with Google. The finding is not flattering to the tool — it says the benefit to junior users may not last — which cuts against the usual worry about company-funded research. AIVoA found no independent replication or outside assessment of the study as of publication.

What the study does establish is that better output and a better worker are two separate measurements, and that it took both. That points employers at the people who are still learning. When a junior lawyer’s assisted work looks as good as a senior colleague’s, the resemblance says little about what the junior can do once the tool is gone.