Lawyers and other expert professionals

Senior lawyers got better at judging work after using AI

A Google-funded study found senior lawyers improved even without AI; juniors did not, on average.
November 14, 2026
Published stamp, level 2 of 5: reputable report, no formal review
The short version, TL;DR
What AI Achieved
In a three-month trial of 133 patent lawyers, AI improved drafts, but after it was removed only seniors got better at spotting errors, on average. Google funded the study. Read more.

A new working paper, which has not been peer-reviewed yet, finds that an AI drafting assistant improved patent lawyers’ work while they used it, but a lasting skill gain showed up mainly for senior lawyers. The three-month experiment was conducted and funded by Google, which built the tool.

Researchers including David Autor and several Google colleagues randomly gave 133 practicing patent lawyers at 11 U.S. firms access to a custom AI patent-writing assistant, later described as part of Gemini Notebook. Blinded expert attorneys scored the work. After 10 days, AI access raised the quality of assisted drafts by 0.34 standard deviations, and after 90 days the assisted gain was 0.38. Junior lawyers also saved time on drafting.

After the AI tool was removed, senior lawyers still did better on an unassisted review task, by 0.45 standard deviations. Junior lawyers showed no average gain.

That unassisted task, redlining a flawed patent application, is a core test of professional judgment. The overall treated group outperformed controls by 0.32 standard deviations, but that advantage was concentrated entirely among seniors, defined here as lawyers with at least seven years of experience.

What happened when the tool was taken away

Junior lawyers’ unassisted scores split: fewer middling results, but more poor scores and more good ones, with no rise in excellent work. The authors suggest the assistant was a springboard for some juniors and a cushion for others. Foundational expertise, they argue, may be needed before AI-assisted practice turns into durable skill.

In follow-up observations, seniors treated the AI more like a “logic auditor,” using it to pressure-test claims and then making deeper edits. Juniors more often stayed with surface changes or diagnosed problems without fully fixing them, a pattern also seen in unassisted juniors.

The study is a preregistered field experiment. Participants included lawyers at firms that already had business relationships with Google. Several authors are Google employees, and funding for the direct costs of the experiment came from Google. David Autor also received support from Google’s visiting fellows program and other foundations.

Why this matters for people

If you hire a lawyer, or if you are becoming one, the people-first question is not only whether AI makes today’s draft look better. It is whether the person across the table is still getting sharper at the judgment that matters when the tool is off, or in front of a client, a colleague, or a court.

This study says the answer may depend on experience. Seniors appeared to harvest more lasting judgment from three months of AI use. Juniors, on average, did not.

Limits to keep in view

This is a working paper, funded and run by the company that made the tool. The sample is modest, the window is only three months, and patent work is a specialized corner of the law. Results may differ in other practice areas, with later models, or over the years it actually takes to become an expert. The authors also note that the lawyers they could recruit were a limited sample of a high-billing profession, and that everyday familiarity with AI has already changed since the trial began.

Statistical “standard deviations” describe average gaps on blinded scores. They are not a promise about every lawyer, every firm, or every future version of the assistant.

Looking ahead for lawyers learning with AI

Better work today and better lawyers tomorrow are not automatically the same thing. In this Google-funded experiment, AI helped drafts immediately, and the lasting skill gain was mainly for senior lawyers, not for juniors on average.

That split is the finding to remember, and it is why this study is more than a simple office productivity success story alone.

You can read the paper as NBER Working Paper 35720. Clients still need both: better papers this week, and lawyers who can catch a damaging error when the assistant is not in the room.

If later, independent studies find the same split, firms may need to be more deliberate about how juniors practice with AI, so today’s shortcut does not get in the way of tomorrow’s judgment.

Like this? Get one every Friday.

The week’s real AI wins, in plain English, each with a stamp.

At a glance

Topic
Who it helps
Lawyers and other expert professionals
Evidence level
Published
Confidence
72
out of 100
How soon
Now
Original Source
NBER Working Paper 35720
Published
September 3, 2026

AI Success Zone summarizes this work in plain English. The original source has the full details and is the final word.

More wins you may like