NBER working paper: RCT of 133 patent attorneys finds AI drafting help boosts junior performance, but only seniors retain benefits without AI
An NBER working paper reports a three-month randomized controlled trial across eleven U.S. IP law firms with 133 patent attorneys. Two-thirds got Google Labs' unreleased AI assistant "InFlow" for two patent claim drafting tasks. Treated attorneys outperformed controls at 10 and 90 days, with larger gains for juniors (under 7 years' experience). In a third task, redlining a patent draft without AI, only senior attorneys kept an advantage. Treated juniors showed no average benefit, with a wider score spread. The study was Google-funded.
Entities: National Bureau of Economic Research (NBER), Google Labs, Google, InFlow, The Decoder, Anthropic
0 primary
What happened
An NBER working paper reports a three-month randomised trial across eleven U.S. intellectual property law firms with 133 patent attorneys. Two-thirds were given Google Labs' unreleased AI assistant, "InFlow", for two patent claim drafting tasks, and they outperformed the control group at both 10 and 90 days, with bigger gains for juniors (under 7 years' experience). In a third task, redlining a patent draft without AI, only senior attorneys kept an advantage. Treated juniors showed no average benefit and a wider spread of scores. Google funded the study.
Why it matters
The finding is a measured data point, not a proof, that AI help can raise junior output without building the underlying skill. Law firms and other professional services firms that rely on juniors learning by drafting should think about how much unaided practice they keep in training. The evidence is narrow: one profession, one tool that the public cannot use, and a working paper that has not been through peer review. It supports caution, not a change of policy.
What is noise
The "quietly stunt their professional growth" framing goes further than the data. The result is an absence of benefit for juniors on one unaided task, not shown harm, and the wider score spread suggests mixed outcomes rather than a uniform decline. Links to the "tragedy of the cognitive commons" and broader talk of expertise erosion are speculation. The coverage also leaves out how large the effects were, and with a small sample and a Google-funded study of Google's own unreleased tool, the results may not carry over to products people can actually buy.
Watch next
- 01Replication or similar trials in other professions or with commercially available tools (for example law, consulting, software), ideally with independent funding and follow-up beyond 90 days.
- 02Peer-reviewed publication of the NBER paper, including effect sizes, confidence intervals and how the redlining task was scored and blinded.
- 03Whether Google releases InFlow, or firms change junior training (for example required unaided drafting) in response, and whether any firm reports data on junior skill development.
Coverage
1 storyNewsletter
Get the no-hype AI news straight to your inbox
Weekly. No spam, unsubscribe any time. Privacy
More labour signals
Full feed →- NYT-led publishers file summary judgment brief citing internal OpenAI/Microsoft messages calling AI training "astonishing theft" and admitting chatbots substitute for journalism18 Sept 202686
- Character.AI files for Chapter 11 bankruptcy and proposes selling assets to Disney as lead bidder for $20 million13 Oct 202677
- Character.AI files for Chapter 11 bankruptcy as part of a deal letting Disney hire most staff and bid for a perpetual license to its technology, leaving liabilities behind19 Oct 202676
- Unsealed filings in NYT v. OpenAI/Microsoft copyright suit reveal internal admissions calling AI training data scraping "theft" and an "existential threat" to publishers17 Sept 202674