NBER trial: AI lifted junior patent attorneys' drafts, but not their unaided review

A working paper from the National Bureau of Economic Research (NBER) reports that AI assistance improved patent attorneys' drafts, most of all for junior lawyers, but that only senior attorneys kept an advantage when the tool was taken away. The Decoder summarised the paper on 24 October 2026. All figures and findings here are The Decoder's account of the authors' work; I have not read the paper.

How the trial was set up

According to The Decoder, 133 patent attorneys across eleven US intellectual property law firms took part in a randomized controlled trial lasting three months. Two-thirds got access to InFlow, an unreleased AI assistant from Google Labs; the rest worked without AI. The study was funded by Google.

There were two claim-drafting tasks, where only the treatment group could use the assistant, and a third task, redlining, where everyone reviewed an existing patent draft without AI. Blinded attorneys at an independent firm scored all results on enforceability, technical accuracy, strategic ambiguity, completeness and clarity.

What it found

• Drafting: the AI group outperformed the control group at both the 10-day and 90-day marks, across all five scoring dimensions. The gain came from fewer bad ratings and more mid-range ones; top scores barely moved.
• Gains were much larger for attorneys with fewer than seven years of experience than for seniors, which matches earlier studies in writing, customer service and software development, per the authors as reported.
• Redlining without AI: the treatment group scored better overall, but the advantage came entirely from senior attorneys. Juniors saw no benefit, even though they had gained most while drafting with AI.
• Among juniors, redlining scores split: fewer mid-range results, more bad ones and more good ones, with no growth at the top.

The authors read the split as AI speeding up skill-building in some juniors and replacing it in others, but The Decoder notes the experiment does not explain why. Those who gained most from the assistant were the same ones who kept the least skill afterwards, which leads the authors to suggest baseline expertise may be a prerequisite for turning AI-assisted practice into lasting ability.

Limits the authors list

• The sample is small, because top attorneys' billable rates made recruiting costly.
• The participating firms are not representative of professional services as a whole.
• There was no baseline skills measurement before the trial.
• Three months is short next to the years it takes to build patent expertise.
• A text analysis found that senior attorneys more often restructured the flawed draft and grounded edits in legal principles, while juniors often spotted serious flaws without fixing them. That pattern also appeared in the control group, so the authors treat it as a baseline gap, not an AI habit.

What it means for a rollout

This is one profession, one funded tool and one short study, so it does not predict your results. What it does offer is a test design you can copy: measure output with the assistant, then measure the same people on a comparable task without it. If your juniors are the ones relying on the tool most, that unaided check is where a problem would show first. Whether the pattern holds for other jobs, or for tools that are on the market, is unknown from this report.