Only seasoned patent lawyers kept AI's skill gains once Autor's trial took the tool away

A law firm pays a junior associate for two things: this week's draft, and the senior lawyer that associate might become. The study set the senior line at seven or more years in practice ^[6]. The tool delivered the first quickly. Drafting scores rose 0.34 standard deviations after 10 days of access and 0.38 after 90 ^[3][2], graded blind by patent attorneys at an independent firm on enforceability, accuracy, strategic ambiguity, completeness and clarity ^[15]. The second showed up, on average, only among lawyers who were already senior ^[4].

"Our results challenge the idea of AI being an automatic skill equalizer," Autor said in written responses to Fortune. "Our data suggest that it's a performance equalizer, but a skill-disequalizer, in that only practitioners who already had foundational mental models were able to level up their underlying skill sets." ^[12]

On a timesheet, faster and better junior drafts ^[11] look like a saving, and a partner could reasonably trim review hours or formal training to match. On this evidence the saving is in output only. The juniors' unaided average stayed flat while the middle of their distribution hollowed out ^[5].

That conclusion can fail in a few ways. The finding may not hold up: the paper is an NBER working paper and has not been peer-reviewed ^[7]. Google built the tool, InFlow, since folded into Gemini Notebook ^[9]. It also conducted the study, paid its direct costs and employs all six of Autor's co-authors ^[13]. MIT's human-subjects committee determined MIT was not engaged in the research ^[13]. Google representatives told Fortune by email that the paper was independent research Autor did as part of his fellowship ^[14]. The sample is small. Two-thirds of 133 lawyers is about 89 with the tool and 44 without ^[16], before any split by seniority, and the published accounts do not give the size of the junior group.

Ninety days may also be too short for a junior to build the mental models Autor describes, and a tool-off test at a year could look different. Or the spread matters more than the average. If a firm can tell early which juniors learn with the tool, AI becomes a way to sort associates, and training money follows the ones who are not learning from it.

We think the trial supports the narrow claim. The tool raises what a junior hands in without, on average, raising what that junior can do alone, so the cost of training juniors stays with the firm. The trial does not show that the tool makes juniors worse on average. A replication outside patent law, or a longer window in which juniors' unaided scores rise, would undercut that view.