AI Made Junior Patent Lawyers Instantly Better. Three Months Later, Only the Seniors Had Kept the Gains

Does AI assistance make professionals better at their craft, or does it quietly replace the craft itself? A new working paper from the National Bureau of Economic Research offers one of the most direct answers yet, and it is uncomfortable: in patent law, at least, the productivity gains were real — but three months later, they belonged almost entirely to the people who needed them least.

The study, "Does AI Assistance Enhance or Erode Expertise?" by MIT's David Autor and six coauthors, was a pre-registered three-month randomized controlled trial across eleven US intellectual property law firms. One hundred thirty-three practicing patent lawyers took part. Two-thirds received access to InFlow, an unreleased AI drafting assistant from Google Labs; the remaining third worked without AI. Google funded the study. All outputs were scored by blinded patent attorneys at an independent firm across five dimensions, including technical accuracy, enforceability and clarity.

While the AI was available, it worked exactly as advertised. The treatment group's benchmark patent drafts scored 0.34 standard deviations above controls at the ten-day mark and 0.38 standard deviations at ninety days, and the biggest beneficiaries were junior lawyers with fewer than seven years of experience — continuing a pattern seen in writing, customer service and software development, where AI narrows the gap between novices and experts.

Then came the twist. In the final task, all participants had to redline an existing patent application with no AI — a core skill patent attorneys routinely perform unaided, and the researchers' proxy for internalized judgment. Treated lawyers still outperformed controls by 0.32 standard deviations, but the entire effect came from senior attorneys, who gained 0.45 standard deviations. Junior lawyers showed no average gain at all. Their score distribution instead split apart: fewer mediocre ratings, but more bad ones and more good ones, with no growth at the top.

The authors read the split as evidence that AI acted as "a springboard for some juniors and a cushion for others." A supplementary text analysis hints at why. Senior attorneys more often restructured flawed drafts and grounded their edits in legal principles, while juniors frequently spotted serious defects without fixing them, leaving margin notes that read like instructions to an assistant. That behavior appeared in the control group too, which the authors interpret as a baseline deficit that AI exposure did not repair. Their conclusion: foundational expertise may be a prerequisite for converting AI-assisted practice into durable skill.

The finding lands in a growing literature on the same worry. An Anthropic study found younger developers who learned a new library with AI help scored 17 percent worse on a later test than those who learned without it; MIT Media Lab researchers documented "cognitive debt" from outsourced essay thinking; and a 2023 BCG experiment showed weaker consultants improved most with GPT-4. Stanford researchers have tied declining entry-level hiring across exposed professions to exactly this substitution of textbook knowledge, and Anthropic recently said junior engineers fell from 21 percent to 14 percent of its engineering staff as AI agents took on more of the work.

The paper's limits are clearly flagged: a small sample driven by senior attorneys' billable rates, firms that may not represent professional services broadly, no baseline skills measurement, and a three-month window that is short against the years it takes to build patent expertise. But the practical stakes are hard to overstate. Even if AI eventually assists live redlining, it will not sit beside a lawyer in a client meeting or a courtroom, where judgment must be available on demand. Firms that measure only the quality of AI-assisted output may be booking productivity gains today at the cost of the next generation's expertise — a trade the authors call a tragedy of the cognitive commons.