Randomized Study: AI Boosted Patent Writing Most for Juniors, but Only Seniors Kept the Gain
A randomized field study across eleven American patent firms found that an unreleased Google assistant made junior attorneys better at writing patent applications – but when the tool was removed, only the senior attorneys retained the advantage.
A three-month randomized controlled study of 133 American patent attorneys suggests that AI assistance can lift immediate performance without building lasting competence – and that experience may determine which of the two effects you get. The results are available as a working paper from the National Bureau of Economic Research (NBER), a US nonprofit research organization, and were reported by The Decoder on October 24, 2026. The study is funded by Google, and the tool tested is an unreleased product from Google Labs.
How the study was set up
The researchers ran a three-month field experiment in eleven American law firms working in intellectual property. 133 patent attorneys were assigned such that two-thirds received access to "InFlow," an unreleased AI assistant from Google Labs, while the final third worked without AI altogether.
All work submitted by participants was assessed by blinded patent attorneys at an independent firm. The assessment covered five dimensions: enforceability, technical accuracy, strategic ambiguity, completeness, and clarity. Effects were measured at two points, ten and 90 days into the study.
Finding one: an immediate writing gain – largest for juniors
On the writing tasks, where the treatment group used InFlow, the result was clear: the treatment group outperformed the control group at both the ten- and 90-day measurements, with effects on all five scoring dimensions.
The strongest effect was found among junior attorneys – those with under seven years of experience. In other words: AI assistance gave the least experienced attorneys the largest immediate improvement in patent writing.
Finding two: without AI, only seniors stood out
Then the result flipped. In a redlining task without AI – designed to measure internalized judgment, meaning what participants had actually absorbed – the treatment group won overall, but the entire advantage came from the senior attorneys. The junior attorneys gained nothing at all, despite having had the largest writing gains during the AI-assisted phase.
The authors read this as a sign that foundational expertise may be a prerequisite for turning AI-assisted practice into lasting competence. They suggest that AI may have accelerated skill-building for some juniors, while for others it functioned as a substitute for it – but the experiment cannot explain why the effects diverged in this way.
Why this matters
The results land in the middle of an ongoing debate about whether AI builds or undermines professional expertise. For law firms and patent writers betting on AI-assisted beginners, the study points to a possible trap: gains measured while the tool is in use may say little about which competence actually sticks once the tool is removed.
At the same time, it is worth keeping perspective. This is a working paper without peer review, funded by Google and based on an unreleased Google product. The study concerns patent attorneys and specifically InFlow, and it has not yet been independently replicated. What conflicts of interest, if any, the working paper documents does not emerge from the available source material.
Open questions
The clearest gap is the mechanism: the study can establish that juniors who benefited most from AI-assisted writing benefited least in lasting skill – but it cannot say why. Without that explanation, it is impossible to know whether the cause lies in how the tool works, how juniors use it, or something else entirely.
Moreover, the central figures – 133 attorneys, eleven firms, the ten- and 90-day measurements – rest on The Decoder's secondary coverage of the working paper, not on the underlying report directly. Until the working paper has been reviewed by other researchers and ideally replicated with a different tool in a different field, the results should be read as a supported hypothesis, not as a firm conclusion about AI and professional practice in general.