News
AI improved patent drafting. The learning gains were less evenly shared.
A patent-attorney study separates better AI-assisted work from unaided skill. AIS explains the findings and what employers should measure.
AI assistance helped patent attorneys produce better drafts, but advantages in later unaided work were not evenly shared. An October 7, 2026 Google Research report described a study in which more-experienced attorneys with prior AI access outperformed the control group on a later unaided task while junior attorneys showed no average advantage.
The result raises a useful distinction for employers: a stronger finished document does not, by itself, establish that its author has developed stronger independent judgment.
Two questions inside one experiment
The researchers' publication abstract describes a three-month randomized trial involving 133 patent lawyers at 11 U.S. intellectual-property firms. It measured drafting with access to an AI assistant and later performance reviewing an existing patent application without AI. Expert graders did not know the participants' treatment assignments.
AI access improved scores on drafting tasks assessed after 10 and 90 days. On the later unaided task, the treatment group's advantage was concentrated among senior attorneys. Junior attorneys had no average gain; their results spread toward both stronger and weaker scores.
What the result does not settle
The authors caution that the sample covers one specialized profession and only three months. The participating firms also did regular, nonexclusive work for Google. Longer-term learning and results in other workplaces remain open questions.
No average gain for juniors is not the same finding as universal deterioration. Nor does this experiment establish that every AI tool, training program or professional task produces the same pattern.
AIS analysis: Measure the worker as well as the work
A polished document can conceal an unfinished understanding. Our takeaway for employers is to evaluate the finished work and the judgment behind it separately.
Faster output answers one management question. Whether someone can recognize a consequential mistake answers another. Ask employees to explain important recommendations, identify what would change their decision and occasionally complete a suitable task without assistance. Review their reasoning alongside quality, time saved and rework.
Make this a learning exercise, not a surprise test of loyalty to a new tool. Give people room to compare their own reasoning with an AI suggestion and discuss why an edit is better or worse. These are practices we recommend, not interventions proven by this experiment.
Keep the tool. Keep the practice, too. A team should be able to challenge an answer before a customer has to do it for them.
Sources and reporting
AIS Empire’s coverage draws on the original reporting and statements below.
- Does better work always mean better workers?
Published 2026-10-07. The October date belongs to this explanatory report, not the date of AIS publication.
- Does AI Assistance Enhance or Erode Expertise? Evidence from a Three-Month Field Experiment in Patent Drafting
Full NBER text was not accessible during this draft review; no full-text-only result is used.
