AI helped patent lawyers write better drafts in a three-month randomized experiment, with the largest estimated gains going to juniors. But when lawyers later reviewed a flawed patent without AI, the average benefit appeared among seniors. Junior lawyers given the assistant performed about the same as juniors without it.
The September working paper, by MIT economist David Autor and six Google-affiliated coauthors, raises a practical question for employers: how much does a better finished document reveal about what its author has learned? The experiment found improved assisted output and uneven evidence of transfer to independent work. It did not establish that junior lawyers lost skills.
Across participants, AI access raised drafting quality by 0.38 standard deviations at roughly 90 days. On the separate, unaided review exercise, the estimated advantage was 0.45 for seniors and -0.03 for juniors. These figures describe differences relative to the spread of control-group scores, not percentage improvements.
The trial randomized 133 practicing patent lawyers at 11 U.S. intellectual-property firms: 90 received access to InFlow, an experimental Google Labs writing assistant, and 43 entered the control group. Assignment accounted for firm and experience. The analysis classified lawyers with fewer than seven years of practice as junior.
Participants completed drafting exercises after about 10 and 90 days. They used invention materials to write dependent claims, which add specific features to a patent’s main claim, and a detailed description. Outside patent attorneys, unaware of participants’ assignments, scored their work for legal strength, technical accuracy, scope, completeness and clarity.
AI access improved drafting scores by 0.34 standard deviations at the first assessment and 0.38 at the second. Junior estimates were larger: 0.58 and 0.60, compared with 0.22 and 0.30 for seniors. The differences between those subgroup effects were statistically uncertain, however. The later drafting exercise was also deliberately harder, so the two assessments do not track improvement on an identical test.
The more revealing exercise asked participants to edit an existing AI-generated patent draft without using AI themselves. They had to repair specified sections, including the claims, and explain the document’s strengths and vulnerabilities. This tested whether experience with the assistant carried over to evaluating and correcting someone else’s work.
Drafting new text and redlining an existing patent are different tasks. The final exercise tested transfer to one related form of professional judgment, rather than whether lawyers retained the same drafting performance. There was no baseline unaided skills test to track each person’s change, and the near-zero junior average does not establish skill loss.
The AI-access group scored 0.32 standard deviations higher overall on redlining. The senior estimate was positive and statistically significant; the junior estimate was close to zero and highly uncertain. Only 24 juniors completed this exercise, compared with 67 seniors. Junior ratings were more spread out, with more poor scores and more relatively good ones, but that pattern cannot identify which individuals learned or regressed.
The seniority distinction also needs care. A significant result for seniors and a nonsignificant result for juniors do not, by themselves, prove that the effects differ. The paper reports directional comparisons but no conventional two-sided interaction estimate with a confidence interval. Its senior-positive, junior-null pattern survived alternative experience cutoffs, supporting the finding without settling its size.
Completion and compliance further limit the conclusion. Ninety-one of the 133 randomized lawyers completed redlining, although the researchers detected no treatment-control difference in attrition. They could not directly enforce the AI ban. After excluding 15 submissions flagged for possible AI use, the senior estimate remained positive but became less statistically certain.
The paper has not been peer-reviewed. Google funded the experiment’s direct costs, the participating firms already did patent work for Google, and Google counsel helped develop the scoring rubric. Random assignment and blinded outside ratings strengthen the comparison within this sample; they do not make it representative of every profession or workplace assistant.
For training, our read is that firms should measure independent judgment alongside assisted output. Having juniors explain and execute corrections themselves is a plausible approach to test. This experiment did not establish a remedy or isolate why the groups differed.
There is evidence elsewhere that interaction design matters. In a peer-reviewed mathematics experiment, high-school students given a direct-answer AI interface improved during practice but scored 17% worse on later unaided exams; a hint-based tutor largely removed that harm. Students learning mathematics are a different population, and those results concern short-term learning. For professional training, the unresolved question is whether a workflow can preserve AI’s drafting gains while requiring juniors to practice the decisions they will eventually own.