NBER study: AI lifts junior patent attorneys' drafts, but only seniors keep the gain
A working paper from the National Bureau of Economic Research (NBER), a nonprofit U.S. economics research organization, asks whether AI assistance builds or erodes professional expertise. It tests the question with a three-month randomized controlled trial of 133 patent attorneys at eleven U.S. intellectual property law firms.
Two-thirds of the participants in the Google-funded study got access to "InFlow," an unreleased AI assistant from Google Labs. The remaining third worked without any AI. The design separated two things: attorneys first worked with AI help, then were assessed without it. Blinded patent attorneys at an independent firm scored all results across five categories: enforceability, technical accuracy, strategic ambiguity, completeness and clarity.
There were three tasks. In two patent claim drafting tasks, only the treatment group could use InFlow, and the control group did the same work unaided. In the third, redlining, all participants had to critically review an existing patent draft without AI. The authors say drafting measures immediate AI-assisted performance, while redlining measures internalized judgment, a task patent attorneys routinely do without any tools.
In drafting, the treatment group clearly outperformed the control group at both the 10-day and 90-day marks, with effects across all five scoring dimensions. The higher scores came mostly from fewer bad ratings and more mid-range ones; top scores barely moved. Gains were much larger for juniors, defined as those with fewer than seven years of experience, than for seniors. The researchers say this fits a growing body of studies showing that AI narrows performance gaps between less and more experienced workers, an effect already seen in writing, customer service and software development.
Redlining produced the opposite picture. The treatment group scored better overall, but the advantage came entirely from senior attorneys. Juniors saw no benefit at all, even though they had posted the biggest gains in AI-assisted drafting. Junior results split apart instead: far fewer mid-range scores, replaced by more bad ones and more good ones, with no growth at the top. The authors read this as AI speeding up skill-building in some juniors while replacing it in others, though the experiment does not explain why the effects diverged. Because the attorneys who gained most from AI were the same ones who retained the least lasting skill, the researchers suggest that baseline expertise may be a prerequisite for turning AI-assisted practice into durable ability.
A supplementary text analysis offers one possible explanation. Senior attorneys more often restructured the flawed draft and grounded their changes in legal principles. Juniors frequently spotted serious flaws without fixing them, leaving notes that read like instructions to an AI assistant. Since this also happened in the control group, the authors do not blame an AI-trained habit. They see a baseline deficit at the junior level that AI use did not fix.
The authors list several limitations: a small sample, driven by top attorneys' high billable rates; participating firms that are not representative of professional services as a whole; no baseline skills measurement; and a three-month window that is short next to the years it takes to build patent expertise. They argue the tension between immediate productivity and lasting skill matters as long as professional work requires judgment without AI, which they expect to be a long time. Even if AI eventually supports redlining, it probably will not be available in real time in client meetings and courtroom proceedings, where judgment calls happen on the spot.
The Decoder places the result alongside similar findings. An Anthropic study found that younger developers who learned a new programming library with AI help scored 17 percent worse on a follow-up knowledge test than those who learned without it. The MIT Media Lab found signs of cognitive debt from outsourced thinking in essay writing. A BCG study in 2023 showed weaker consultants improving far more with GPT-4 than strong performers. Stanford researchers explain the decline in entry-level jobs across industries by arguing that AI can mostly replace textbook knowledge, the main thing early-career workers bring, while experienced professionals also carry hands-on expertise AI cannot easily substitute. The article also notes that Anthropic recently said junior positions in its engineering departments fell from 21 percent to 14 percent since early 2025.
The piece frames the study as more experimental evidence for what some researchers call the "tragedy of the cognitive commons": individual firms benefit from automation while the cost of stunted talent development spreads across a whole profession.
Key facts
- An NBER working paper reports a three-month randomized controlled trial with 133 patent attorneys at eleven U.S. IP law firms; two-thirds of participants got Google Labs' unreleased InFlow assistant, the rest worked without AI.
- In drafting, the AI group beat the control group at both the 10-day and 90-day marks across all five scoring dimensions, with much larger gains for juniors (fewer than seven years of experience) than for seniors.
- In an unaided redlining task, the treatment group scored better overall, but the advantage came entirely from seniors; juniors saw no benefit, and their scores split into more bad and more good ones.
- The authors suggest baseline expertise may be a prerequisite for turning AI-assisted practice into durable ability, but the experiment does not explain why junior effects diverged.
- Limitations include a small sample, unrepresentative firms, no baseline skills measurement and a three-month window; the study was funded by Google.
Why it matters
The study separates two things that are usually blurred: how good AI-assisted work is, and what the worker can still do once the AI is gone. The people who gained the most from AI in drafting, the juniors, were the ones who kept the least in the unaided review task. That puts a specific, experimental data point behind the worry that AI helps the least experienced do better work today while their own skill grows less. The Decoder links it to what some researchers call the "tragedy of the cognitive commons": firms gain from automation, while the cost of stunted talent development lands on the whole profession.
Who it affects
Most directly, junior patent attorneys, defined here as those with fewer than seven years of experience, and the firms that train them. Senior attorneys kept a benefit in the unaided task. The same pattern of AI narrowing gaps between less and more experienced workers has been seen in writing, customer service, software development and consulting, so the question reaches other professions where judgment must be exercised without tools. The article also mentions Anthropic's statement that junior roles in its engineering departments fell from 21 percent to 14 percent since early 2025, and Stanford researchers' argument about entry-level jobs.
How to use it
There is nothing to try yet: InFlow is an unreleased Google Labs assistant. What carries over is the study design and the authors' framing. They argue that as long as professional work requires judgment without AI, immediate productivity gains and lasting skill development pull in different directions, and that AI probably will not be on hand in real time in client meetings and courtrooms. Seniors who restructured flawed drafts and grounded changes in legal principles kept the benefit; juniors who flagged flaws without fixing them did not.
How solid is it
It is a working paper, not a peer-reviewed publication as far as the article says, built on a randomized controlled trial. Results were scored by blinded patent attorneys at an independent firm across five categories, which strengthens the measurement. The study was Google-funded and used a Google Labs tool. The drafting result is clear across both time marks and all five dimensions. The redlining result is reported as a pattern, and the article gives no effect sizes (score points or percentages) for either task. This account rests on The Decoder's report of the paper.
Risks and caveats
The authors themselves list a small sample driven by top attorneys' high billable rates, firms that are not representative of professional services as a whole, no baseline skills measurement, and a three-month window that is short compared with the years it takes to build patent expertise. The experiment does not explain why junior effects diverged. The authors read the split as AI speeding up skill-building in some juniors and replacing it in others. The text analysis found that the junior habit of spotting flaws without fixing them also appeared in the control group, so the authors treat it as a baseline deficit rather than something AI taught. The article does not claim the study proves AI causes skill erosion; the authors only suggest baseline expertise may be a prerequisite for durable gains.
“tragedy of the cognitive commons”
— phrase used by some researchers for the erosion of professional expertise, as cited by The Decoder