The hopeful story about AI at work goes like this: hand a junior a powerful tool, and they produce work like a veteran. A new study of patent lawyers suggests the tool does something more unsettling. It makes the junior's output better while leaving the junior no better. David Autor, the MIT economist best known for documenting the "China shock" of lost manufacturing jobs, put it bluntly to Fortune: AI is "a performance equalizer, but a skill-disequalizer."
The experiment, a National Bureau of Economic Research working paper not yet peer-reviewed, enrolled 133 lawyers at 11 US intellectual-property firms. Two-thirds got access to an unreleased Google patent-drafting assistant; the rest did not. They drafted patents at day 10 and day 90, and independent attorneys graded the work blind. After 90 days, AI access lifted drafting scores by 0.38 standard deviations, about 11 percentile points. Crucially, that gain came from fewer weak drafts, not more brilliant ones. The floor rose; the ceiling did not.
Then came the real test. Lawyers had to mark up a flawed patent with no AI allowed, a task the paper calls one they "routinely perform unaided." Here the two groups split. Senior lawyers who had used AI beat their peers by 0.45 standard deviations. Junior lawyers showed no average gain at all. Three months with a capable assistant had sharpened the people who already knew the law and left the novices roughly where they started.
Autor's explanation turns on what each group did with the tool. The veterans treated it as a "logic auditor," using it to pressure-test structure while they focused on strategic scope, the forest rather than the trees. But, as he notes, "you can use AI as a 'logic auditor' only if you already know the law well enough to spot when fluent text is legally flawed." The juniors, lacking that foundation, polished introductory prose, flagged serious flaws in comments instead of fixing them, and mistook a smoother workflow for progress. Their biggest measured gain was in job satisfaction: they loved skipping the blank page and stepping straight into a reviewer's chair.
That satisfaction is the trap. "Young professionals need to beware of the illusion of competence," Autor warns. "The only way you're really going to know if you're developing skills is if you do tasks without AI assistance and evaluate your performance." The fluency the tool provides can feel like mastery while the underlying mastery never forms, a gap that stays hidden until the AI is taken away.
The paper carries real caveats. The sample was small, drawn from firms already working for a single sophisticated client, and three months is nothing against the years it takes to build patent expertise. Only 91 of 133 finished the unassisted test, and some of the senior advantage weakens under stricter statistical thresholds. Autor is careful to say this does not prove AI stops juniors from ever becoming experts. It is a snapshot, not a verdict.
But the warning for employers is sharp enough. The temptation to hire fewer juniors because the output looks fine is, Autor argues, short-sighted: "If firms automate away formative practice without replacing it with some kind of guided learning environment, they sever the apprenticeship pipeline that produces tomorrow's senior partners." His deeper point echoes his China work. Real advantage has always lived less in tools than in hard-won human know-how. "If we think that AI will relieve us of the burden of mastering expertise," he said, "I think we will be sorely disappointed."