AI boosts output but splits junior skill growth in patent law study

Google Research conducted a three-month randomized controlled trial with 133 patent attorneys to measure how AI assistance affects both immediate work quality and long-term skill acquisition. The study, published in October 2026, reveals that while AI tools consistently improve drafting performance across experience levels, their impact on developing independent professional judgment varies significantly between senior and junior lawyers.

What happened

The experiment involved eleven intellectual property law firms that maintain regular business relationships with Google. Researchers randomized access to an unreleased Google Labs AI patent writing assistant, now integrated into Gemini Notebook, among the participating lawyers. Two-thirds of the attorneys received early access to the tool as the treatment group, while the remaining third received training on AI usage but were withheld from using the specific tool until after the study period ended.

Participants completed two drafting tasks using simulated inventor materials: one after ten days of AI access and another after ninety days. Independent legal experts graded these drafts on five dimensions: enforceability, accuracy, strategic ambiguity, completeness, and clarity. At the ninety-day mark, researchers added a critical redlining task where lawyers had to manually correct errors in a hypothetical patent without any AI assistance. This unassisted task served as the primary measure of whether AI usage had translated into durable skill development or merely temporary output enhancement.

How it works

The study distinguishes between using AI as a productivity booster and using it as a learning scaffold. For senior lawyers, the AI tool functioned as a "logic auditor" rather than a final answer generator. These experienced professionals used the AI output to challenge their own assumptions, forcing them to articulate the reasoning behind structural edits and legal doctrines. This process strengthened their foundational expertise by weakening their attachment to existing prose and encouraging deeper engagement with the logical structure of patent claims.

In contrast, junior lawyers often treated the AI output as a finished product or focused on surface-level improvements. Their editing patterns showed rigid formalism, such as working sequentially from top to bottom and spending excessive time on low-stakes introductory sections. When errors were identified, juniors frequently left diagnostic comments rather than executing the necessary fixes. This "diagnose-without-execute" behavior persisted regardless of AI access, suggesting that the tool did not remediate the baseline deficit in practical execution skills that characterizes early-career professionals.

Key details

• AI access improved drafting scores by 0.34 standard deviations at ten days and 0.38 standard deviations at ninety days, equivalent to an 11-percentile point increase.
• Senior lawyers with seven or more years of experience outperformed controls by 0.45 standard deviations on the unassisted redlining task, indicating significant judgment gains.
• Junior lawyers showed no average improvement in unassisted judgment; instead, their scores bifurcated into more very low and more good scores, with no increase in excellent scores.
• Juniors using AI completed drafting tasks 18 minutes faster than the control group’s average of 124 minutes, highlighting immediate efficiency gains despite mixed learning outcomes.
• The study involved 133 lawyers across eleven firms, with blinded grading by third-party patent professionals to ensure objective evaluation of quality dimensions.

Why it matters

For engineering leaders and technical founders, this study highlights a critical tension between short-term productivity and long-term capability building. While AI tools can immediately elevate the quality of output and reduce time spent on routine tasks, they may inadvertently short-circuit the tedious but formative experiences required to build deep expertise. If junior team members rely on AI to handle complex reasoning tasks without engaging in the underlying logic, they may fail to develop the judgment necessary for high-stakes decision-making when tools are unavailable or inappropriate.

The findings suggest that simply providing AI access is insufficient for holistic team development. Organizations must actively design workflows that prevent AI from becoming a crutch that masks skill gaps. Without intentional intervention, companies risk creating a workforce that is highly productive in assisted environments but lacks the resilient, independent judgment needed for novel problems, client interactions, and strategic oversight. This polarization in junior performance indicates that AI can amplify existing differences in approach rather than uniformly elevating baseline competence.

What you can do

• Implement mandatory unassisted review phases where junior engineers or analysts must critique and correct AI-generated outputs without tool support to practice execution skills.
• Structure mentorship programs that require seniors to explain the "why" behind AI suggestions, turning the tool into a discussion prompt rather than a solution engine.
• Monitor performance distributions for polarization, looking for increases in both high and low extremes among junior staff as a sign that foundational skills are not being evenly reinforced.
• Design tasks that force engagement with core logic, such as requiring manual reconstruction of key components from scratch, to prevent over-reliance on surface-level AI edits.
• Evaluate hiring and promotion criteria to value independent judgment and problem-solving under constraints, ensuring that AI-assisted productivity does not become the sole metric of success.