On October 7, 2026, Google Research published "Does better work always mean better workers?" by David Autor and Tanya Rodchenko. It summarises a pre-registered field experiment written up as NBER Working Paper 35720, "Does AI Assistance Enhance or Erode Expertise?" The question is one most AI productivity studies skip: does better output with AI mean the person is building skill?
The design
The team randomised access to a then-unreleased Google Labs patent-drafting assistant, now part of Gemini Notebook, among 133 lawyers at eleven IP law firms that do regular, non-exclusive business with Google. Two-thirds of lawyers at each firm got the tool, and the rest got some AI training but no access until afterwards. Independent experts, blind to condition, graded drafting tasks at 10 and 90 days. At 90 days everyone also redlined a flawed patent application without AI, which measured unassisted skill.
The results
- Assisted drafting improved for everyone. Scores rose by 0.34 standard deviations at 10 days, about 10 percentile points, and by 0.38 SD at 90 days. The authors say most of the gain came from fewer poor drafts, not more excellent ones.
- Juniors also got faster. On the 10-day task they were 18 minutes faster than the control group's average of 124 minutes.
- The unassisted test split by seniority. Treated lawyers scored higher overall, but the authors write that the effect "was driven entirely by senior lawyers, who outperformed control subjects by 0.45 SD (a 13-point climb), while juniors showed no discernible improvement." Junior scores "bifurcated: more very low scores, fewer mediocre ones, more good ones, and no more excellent ones."
The NBER paper concludes that foundational expertise may be a precondition for turning AI-assisted practice into lasting skill.
How it fits with last week's study
This is a different finding from the COLM study covered here on October 11, where ten minutes of AI help left people worse and quicker to give up once it was removed. Autor's trial ran three months, with real professionals doing their real job, and found no average harm. Seniors gained judgement. Juniors' assisted work improved, but that did not reliably turn into faster skill. The two studies agree on the part that matters for managers: the people with the least expertise are the least likely to learn from the tool.
What remains uncertain. The authors note the small sample, the focus on top-tier, high-billing professionals, and a three-month window far shorter than it takes to build expertise. The models have also improved since the trial ran. Google built the tool and ran the study, and the firms do business with Google, though the design was pre-registered and grading was blind.
Analysis: For engineering leaders, the parallel to junior developers using coding agents is direct. Tool access can raise the floor on today's output without building the judgement those people will need to review agent output later. The design response is to keep some unassisted practice in the loop: reviews where juniors explain the change, rotations without the agent, or tasks graded on reasoning rather than the diff. Seniors appear to learn from AI on their own. Juniors need the learning designed in.
In a pre-registered three-month RCT with 133 patent lawyers, AI drafting help raised quality by 0.34-0.38 SD for everyone, but on a later unassisted task only senior lawyers improved (0.45 SD) while junior scores split toward both poor and good, suggesting foundational expertise is needed to learn from AI-assisted work.