MIT's Autor: AI Boosts Output but Widens the Skills Gap

AI access raised patent-drafting performance by 0.38 standard deviations over 90 days — but when researchers took the tool away, only lawyers with seven or more years of experience showed a genuine skill gain. Juniors showed none. That is the central finding of a National Bureau of Economic Research working paper by MIT economics department head David Autor and six Google-employed co-authors, and it inverts the popular promise that AI levels the playing field for young workers.

"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."

The study sorted lawyers by experience, not age — "senior" meant seven-plus years in practice, and the researchers did not study Gen Z as a group. But early-career workers are the people the findings bear on most directly.

Who ran the experiment, and how?

The paper, which has not been peer-reviewed, enrolled 133 lawyers at 11 U.S. intellectual-property law firms with ongoing patent-drafting relationships with Google. Researchers randomly assigned access to a custom AI patent-drafting assistant — a then-unreleased Google Labs tool. Two-thirds of the lawyers got access. The rest went without until the study ended.

Google's role went beyond the tool. The paper lists Autor's six co-authors as Google employees and states that Google conducted the study and paid its direct costs; MIT's human-subjects committee determined MIT was not engaged in the research. Google representatives described the paper in an email to Fortune as independent research Autor did as part of his fellowship.

Lawyers drafted patents from simulated inventor materials after 10 days and again after 90 days. Patent attorneys at an independent law firm graded the work blind, scoring enforceability, accuracy, strategic ambiguity, completeness and clarity.

What did the numbers show?

After 90 days, AI access raised drafting scores by 0.38 standard deviations. In a Google blog post accompanying the paper, Autor and co-author Tanya Rodchenko described that as an 11-percentile-point gain relative to control scores. The improvement came from less weak work, not more excellent work.

The harder test came at the end. Lawyers marked up a hypothetical patent containing many substantive and stylistic errors — a task patent lawyers "routinely perform unaided." No AI was allowed. Senior lawyers who had AI access beat their control-group peers by 0.45 standard deviations. Junior lawyers showed no average gain.

The caveats are material:

• Only 91 of the 133 lawyers finished the AI-free test.
• The firms declined a baseline skills test, so the researchers inferred skill gains from random assignment rather than before-and-after measurement.
• The authors flagged 15 of the 91 markups as possibly AI-assisted despite the ban, about as often among control lawyers. Excluding them, the senior advantage shrinks to 0.39 standard deviations and holds only at a looser statistical threshold. The overall effect does not.

What happened to the juniors?

The juniors' scores split: significantly more poor scores, fewer mediocre ones, more good ones and no gain at the top. The authors write that AI "served as a springboard for some juniors and a cushion for others."

The juniors tended to work top to bottom, polishing introductory text before reaching the main claims. Some spotted serious flaws but left comments describing them instead of fixing them — a pattern the authors call "a baseline junior deficit" that three months of AI access did not fix.

The juniors did notice something: they liked the tool. Autor said AI-assisted junior lawyers reported a surge in task satisfaction, the largest effect in the study. They felt that skipping the "blank page problem" let them step straight into a reviewer role.

"Young professionals need to beware of the illusion of competence," he said. "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."

Why did the seniors gain?

The experienced lawyers described AI in follow-up interviews as a "logic auditor," not a finished product. The authors write that it weakened their attachment to existing prose and forced them to spell out the why and how of their structural edits. On the unassisted test, treated seniors skipped low-stakes prose, rebuilt claims from scratch, cut language that could narrow legal protection and tied many edits to legal doctrine.

"Arguably, 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," Autor told Fortune. He said offloading baseline drafting freed seniors to focus on strategic scope — "the forest not the trees."

His advice to novices: work like seniors. Frame the argument unassisted first, then use AI as a critic. He acknowledged that is a lot to ask. "Most of us don't have the self-discipline to do things by hand when a readily available, low-effort tool is sitting right there, offering to do it for us," he said.

What should employers do?

For employers, better output creates a temptation to hire fewer juniors. Autor called that 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," he told Fortune.

He suggested firms separate tasks that AI can permanently complete from judgment tasks requiring unassisted mastery, pair AI rollouts with regular unassisted skill checks such as offline redlining exercises, and hold partners accountable for mentoring juniors through the technology's mistakes. These are ideas, not proven fixes. "It's early days in this area of research and practice," Autor said.

The limits are real. The sample was small, drawn from firms doing patent work for a single sophisticated client, and three months is brief against the years it takes to build patent expertise. The blog post notes that newer models and wider AI familiarity might change the results today. The study does not show that AI keeps juniors from becoming experts over a career.

Autor, best known for the "China shock" research with David Dorn and Gordon Hanson showing Chinese import competition depressed wages and employment in exposed U.S. labor markets for at least a decade, has pushed back on comparisons between AI and that shock. On the Possible podcast hosted by LinkedIn co-founder Reid Hoffman, he said AI "will not be, in any sense, a repeat of the China trade shock" because AI will have a "very different texture" in boosting productivity. His stated worry is not mass job loss but churn: AI "eliminates some areas of specialty while creating new ones simultaneously," and the workers losing old careers are generally not the ones who can take the new roles.

Most technological advantage, Autor told Fortune, comes from human know-how built through slow, laborious mastery. "If we think that AI will relieve us of the burden of mastering expertise, I think we will be sorely disappointed," he said. "Human intelligence and machine intelligence will be complements for the long term."