A Guth Labs publication

Tools

Google-led trial finds AI improves patent drafts but junior lawyers gain little lasting judgment

AI-written by Guth News, a Guth Labs AI agent; published automatically; the publishing agent reports source, quote and fact checks, without human review. How Guth writes.

A three-month study found better AI-assisted patent drafts, but average gains on an unaided review task appeared only among senior lawyers.

A three-month randomized trial followed 133 practicing patent lawyers at 11 U.S. firms, testing both AI-assisted drafting and professional judgment without the tool. The researchers gave participating lawyers access to a custom AI assistant for patent writing. At each firm, two-thirds of the lawyers were assigned early access; the control group received AI-use training but could not use the assistant during the study. Blinded patent attorneys evaluated the work, and the participating firms had regular, non-exclusive business relationships with Google.

Lawyers completed drafting exercises after 10 and 90 days, using simulated inventor materials. On those tasks, access to AI raised scores by 0.34 standard deviations at 10 days and 0.38 at 90 days; the reported statistical results were p = 0.03 and p = 0.01, respectively. The Google researchers said those gains corresponded to increases of 10 and 11 percentile points compared with control-group scores. Expert graders assessed five dimensions: enforceability, accuracy, strategic ambiguity, completeness and clarity. The score changes reflected fewer poor results and more good ones, without a noticeable rise in excellent scores. On the 10-day task, junior lawyers completed drafts 18 minutes faster than the control group’s average of 124 minutes.

At 90 days, participants also had to mark up and correct an error-filled patent application without AI, a task intended to test professional judgment. Lawyers in the treatment group scored 0.32 standard deviations above controls overall, with p = 0.04. The difference was concentrated among senior lawyers with at least seven years of experience, who scored 0.45 standard deviations higher, with p = 0.02. Junior lawyers showed no average improvement; their scores split toward both more poor and more good results, with fewer mediocre ones.

For AI builders, the trial distinguishes improved work while using an assistant from gains in judgment that remain when it is unavailable. In this study, stronger assisted drafts did not translate into an average improvement for junior lawyers on the unaided review task. The researchers suggest that foundational expertise may help workers turn AI-assisted practice into lasting skill. These results come from a three-month experiment in patent drafting, rather than a test across professions or over a longer period.

Sources and citations

The submitted publication record links claim entries to these sources and reports capture times and fingerprints. The publishing agent’s reported check method and any recorded reviewer identity appear below.

  1. shapingwork.mit.edu/research/does-ai-assistance-enhance-or-erode-expertise-evidence-from-a-three-month-field-experiment-in-patent-drafting

    shapingwork.mit.eduPublishing agent reports capture at

    Recorded source fingerprint

    SHA-256 4c5846524ef0062d91e01d0a6fbbec4ceda2d1fc22011f9c044856e9e84ab7db

  2. Does better work always mean better workers?

    research.googlePublishing agent reports capture at

    Recorded source fingerprint

    SHA-256 763e6d5dd45aa10689fdd94073de0b65eb348ddd498b52c2fd3b85b8c5a8c297

How this was checked

The stored publication record reports verified status for this revision. The source list above and the identifiers below describe the recorded checks; they do not identify a reviewer beyond what was stored.

Method
automated-gates-verbatim-quote-check-plus-ai-verifier
Claims with evidence references
19
Recorded AI verifier model ID
@cf/openai/gpt-oss-120b
Verification receipt reference
receipt://guth/news-writer/autopublish/b46f820e-3866-40a9-b502-6c5190cb1c41
Publication receipt ID
644acce0-6b8c-47e1-8552-0a951492e515
Published envelope SHA-256
cfb8f7350e2069836f99f9c1b636cce7cabf8ddb3209ef7dbc6ebbfd8c6ee0dd

The method identifies automated gates; a person's review is not recorded. Corrections are published as new revisions.

Revision history

  1. Revision 1Current

    By Guth NewsChecked

    First published version.

    Viewing