Most studies of AI at work measure the work. This one also measured the worker, three months later, with the tool taken away. The two measurements point in different directions.
David Autor of MIT and six Google-affiliated researchers ran a pre-registered randomized controlled trial across eleven US intellectual property firms (NBER Working Paper 35720, September 2026, revised October 2026). Researchers randomized 133 practicing patent lawyers across 11 US firms; 91 completed the three-month study protocol. Lawyers were assigned two to one, stratified by firm and experience, to receive a custom AI patent drafting assistant or to work without one for three months. Every piece of work was graded by blinded expert patent attorneys at an independent law firm, against a rubric covering enforceability, accuracy, strategic ambiguity, completeness and clarity.
This article examines a September 2026 NBER working paper revised in October, following renewed discussion on October 25.
The output result is the familiar one
With AI, the drafts got better. The treatment group outscored controls by 0.34 standard deviations after 10 days and 0.38 after 90 days, across all five scoring domains, with larger point estimates for junior lawyers than for seniors. That fits the pattern reported in writing, customer support and coding studies: the tool helps the least experienced most.
One sentence in the paper deserves more attention than the headline effect. The authors note that the higher average scores reflect a lower prevalence of weak work, not a higher prevalence of excellent work. AI, in this trial, raised the floor. It did not raise the ceiling. In this experiment, AI assistance mainly reduced low-quality drafting outcomes; the researchers did not observe a corresponding increase in exceptional scores.
The time saving was real but modest, and it belongs to the assisted work only. On the 10-day drafting task, treated lawyers finished an estimated 10 minutes faster than controls (p = 0.05), against a control-group mean of 110 minutes; the 90-day estimate was similar in size but less precise. On the final unassisted task, there was no time saving at all.
The skill result is not the familiar one
At the end of the three months, every participant redlined an existing patent application without AI, a routine task chosen because it tests judgment rather than text production. The clearest result belongs to senior lawyers, defined as seven or more years of experience, who gained 0.45 standard deviations (p = 0.02). The pooled treatment advantage was 0.32 standard deviations (p = 0.04). The overall unassisted advantage was sensitive to excluding submissions flagged for possible AI use; the senior subgroup effect was more persistent, though its significance weakened under that restriction.
Junior lawyers showed no average gain on the unassisted redlining task. Their score distribution shifted away from the middle, with more poor and more good results, consistent with divergent experiences but not proof that AI caused identifiable individuals to learn or lose skill. The paper's quintile breakdown puts numbers on the split: significantly fewer mediocre scores (a fall of 0.42 in share, p = 0.01), significantly more poor scores (a rise of 0.22, p = 0.04), and a rise in good scores that is imprecisely estimated (p = 0.20), with no gain at the top of the distribution. The authors' reading is that the assistant acted as a springboard for some juniors and a cushion for others, and they are explicit that the experiment cannot say which juniors got which effect, or why.
The disclosure and the caveats
The experiment was conducted and funded by Google. Six of the seven authors were Google-affiliated, including a disclosed contractor; David Autor participated through Google's visiting-fellows program. All eleven participating firms had existing relationships with Google. Google's involvement and the participating firms' existing relationships limit how confidently the findings can be generalized to other tools, workplaces and training arrangements. The study's blinded independent grading is an important design strength, but it does not eliminate those limitations. The sample, by the authors' own account, best represents firms already sophisticated in patent practice.
Two further cautions belong on the record. The tasks were simulated, built from hypothetical inventor materials, in a single profession, with a single custom tool. And the estimates rest on the lawyers who completed each task — 105 of the 133 finished the 10-day task, 98 the 90-day drafting task and 91 the full protocol — though the paper finds no evidence that attrition differed by treatment arm. The unassisted task also carried a contamination risk the authors checked themselves: audits flagged 15 of the 91 redlining submissions where AI use appeared possible. Those submissions scored 0.18 SD higher than unflagged ones, a statistically insignificant difference (p = 0.31). Excluding them drops the pooled unassisted effect from 0.32 to 0.25 SD, below conventional significance (p = 0.14), and the senior effect from 0.45 to 0.39 SD, which remains significant at the 10 percent level (p = 0.06). With LLM ratings in place of human ones, both estimates are larger and significant at the 1 percent level. The senior result is sturdier than the headline average; quote them in that order.
What a firm should do with this
The practical reading is not that juniors should be kept away from AI. Their assisted output improved the most, and clients are not paying for unaided exercises. The reading is that assisted output and trainee development are separate line items, and this trial is randomized evidence that buying the first does not automatically deliver the second. The authors suggest that existing expertise may help experienced lawyers learn from AI-assisted work. The trial did not establish why senior and junior outcomes differed. Nor did it test the arrangements a firm might pair with the tool — mentoring, review, structured practice, assessment — so it cannot say whether those close the gap. What it establishes is narrower: output and skill are separate measurements, and only one of them was guaranteed to move.
Sources
- [1] Autor, Rodchenko, Martin, Iscenko, Strand, Pearl and Ferere, "Does AI Assistance Enhance or Erode Expertise? Evidence from a Three-Month Field Experiment in Patent Drafting", NBER Working Paper 35720 (September 2026, revised October 2026) — abstract, design, all effect sizes, author disclosure statements and the flagged-submission robustness checkRead source
- [2] MIT Stone Center on Inequality and Shaping the Future of Work, working paper page for the same study (September 2026) — abstract and summaryRead source
- [3] Paper PDF via MIT Economics (dated October 7, 2026) — introduction, methods and appendix: 2:1 randomization, the five-domain rubric, at least two raters per task, task completion counts, the seven-year seniority cutoff and its robustness at 3–11 years, the time-on-task results, and the flagged-submission analysis (Tables A4–A5)Read source