AI can help patent attorneys produce stronger drafts, with the largest immediate gains among less experienced lawyers. A three-month field experiment suggests those gains may not translate into lasting skill for everyone: in a later task without AI, the overall advantage appeared among senior attorneys, while junior attorneys showed no benefit.
How the experiment separated AI help from independent judgment
The randomized controlled trial included 133 patent attorneys from eleven U.S. intellectual property law firms. Two-thirds received access to InFlow, an unreleased AI assistant from Google Labs, and the remaining third worked without AI. The study was funded by Google.
Participants completed two patent claim drafting tasks. At the 10-day and 90-day marks, only the treatment group could use InFlow; the control group did the same work unaided. The researchers compared the quality of those drafts with a separate exercise designed to assess what attorneys could do on their own.
In that third task, called redlining, everyone reviewed an existing patent draft without AI. Blinded patent attorneys at an independent firm scored the work for enforceability, technical accuracy, strategic ambiguity, completeness, and clarity. Drafting measured performance with assistance; redlining tested internalized judgment in work attorneys routinely do without tools.
AI narrowed the gap in drafting performance
Attorneys with AI access scored better on the drafting tasks at both measurement points. Improvements appeared across all five scoring categories. The change came mainly from fewer poor ratings and more mid-range scores, while the number of top scores barely changed.
The gains were considerably larger for juniors with fewer than seven years of experience than for senior attorneys. In other words, AI raised the quality of junior work more sharply, narrowing the performance difference between less and more experienced lawyers on the assisted tasks.
The researchers say this pattern is consistent with findings in other areas, including writing, customer service, and software development, where AI has helped less experienced workers improve their performance. But a better output while AI is available does not, by itself, show that the user has learned how to produce that work independently.
Independent review showed a different pattern
In redlining, the treatment group performed better overall, but the advantage came entirely from senior attorneys. Junior attorneys who had made the biggest gains in AI-assisted drafting showed no benefit in this unaided review.
Junior results were mixed. They had fewer mid-range scores and more low and high scores, but no increase in top scores. The researchers suggest that AI may have accelerated learning for some juniors while substituting for practice for others. The experiment did not establish why the results diverged.
Those who gained most from AI during drafting were also the group that retained the least lasting skill. The authors propose that a foundation of expertise may help workers turn AI-supported practice into durable ability, though the study cannot prove that explanation.
A supplementary analysis of the written reviews offers a possible clue. Senior attorneys more often rebuilt the flawed draft and tied their edits to legal principles. Juniors often identified serious problems but left notes describing what should be fixed instead of making the changes themselves. That pattern also appeared among juniors in the control group, so the authors treated it as an existing gap that AI use did not resolve, rather than a habit caused by AI.
Why the findings matter, and what they cannot show
For firms, AI assistance can improve work in the moment, especially for junior attorneys. The findings also raise a question for professional development: if AI does more of the drafting, do early-career workers get enough practice making and applying the judgments that experienced colleagues have built over time?
The distinction matters because professional work sometimes requires decisions without AI. The researchers argue that this tension between productivity and skill development is likely to persist. Even if AI supports redlining, it may not be available in real time during client meetings or courtroom proceedings, where judgment calls happen on the spot.
The trial has clear limits. The sample was small, partly because top attorneys have high billable rates, and the participating firms may not represent professional services broadly. The study also lacked a baseline measurement of participants' skills, and its three-month window was short compared with the years needed to build patent expertise.
These results add experimental evidence to a wider debate about whether automation can weaken the development of expertise across a profession. Individual firms may gain from AI-supported work while the longer-term cost of fewer opportunities to learn could reach beyond any one workplace. The study points to a risk worth examining, while leaving open how common or lasting that effect may be.