Gen Z was promised that artificial intelligence would be a great equalizer: a tool that could allow a 24-year-old to produce work comparable to that of a veteran. New research by economist David Autor suggests AI may instead narrow differences in immediate performance while widening gaps in underlying skills.

In a three-month experiment involving patent lawyers, AI improved the quality of drafts across experience levels. But when researchers removed the tool and tested the lawyers' independent judgment, only experienced lawyers showed an average improvement. Junior lawyers did not.

“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 grouped lawyers by experience rather than age. “Senior” lawyers had at least seven years in practice, and the researchers did not examine Gen Z as a separate group. However, the findings are most directly relevant to early-career workers who are still developing professional expertise.

From the China shock to AI

Autor is head of MIT’s economics department and a Google Technology and Society Visiting Fellow. He is best known for the “China shock,” a body of research developed with economists David Dorn and Gordon Hanson. Their work found that Chinese import competition caused lasting harm to American manufacturing communities, which did not adjust as smoothly as optimistic accounts of globalization had predicted.

In heavily exposed labor markets, wages and labor-force participation remained depressed while unemployment stayed elevated for at least a decade. More recently, Autor and Hanson have warned of a “China shock 2.0,” referring to Chinese competition in advanced industries that could threaten American technological leadership and the high-wage jobs associated with it.

Some prominent economists, including Torsten Slok of Apollo Global Management, have said an “AI shock” could have similarities to the China shock in its effect on the U.S. labor force. Bhaskar Chakravorti of Tufts has coined the phrase “the wired belt” as a parallel to the Rust Belt.

Autor rejected that comparison during an appearance 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 the China shock was experienced by U.S. firms as a “pure negative competitive shock,” while AI would have a “very different texture” by increasing productivity.

The debate over AI rests partly on an assumption that the China shock challenged: once technology takes over routine work, workers will naturally advance to higher-level skills. Autor told Fortune that his concern about AI is not primarily mass job loss.

“The danger that we may face with AI is not that it eliminates large numbers of jobs in net, but that it eliminates some areas of specialty while creating new ones simultaneously,” he said. “That sounds like a reasonable deal, but the workers losing those old careers are generally not the ones who will be able to take advantage of the new roles that emerge.”

The patent-law experiment examined a narrower question: whether workers using a powerful new tool are actually developing the expertise required to advance in their field.

Better drafts, uneven learning

Autor and six co-authors published the study as a National Bureau of Economic Research working paper. The paper has not been peer-reviewed — working papers are circulated for discussion and comment before any journal review.

The research enrolled 133 lawyers at 11 U.S. intellectual-property law firms with ongoing patent-drafting relationships with Google. Researchers randomly assigned the lawyers access to a custom AI patent-drafting assistant, which was then an unreleased Google Labs tool. Two-thirds received access, while the remainder did not use the tool during the study. That randomization is central to the design: because access was decided by chance, differences in outcomes between the two groups can be traced to the tool rather than to pre-existing differences in ability.

Google’s role extended beyond providing the software. The paper identifies Autor’s six co-authors as Google employees. It also says Google conducted the study and paid its direct costs, while MIT’s human-subjects committee determined that MIT was not engaged in the research. In an email to Fortune, Google representatives described the paper as independent research Autor conducted as part of his fellowship.

The lawyers drafted patents from simulated inventor materials after 10 days and again after 90 days. According to the paper, patent attorneys at an independent law firm graded the work without knowing which lawyers had used AI. They evaluated enforceability, accuracy, strategic ambiguity, completeness and clarity.

After 90 days, access to AI increased drafting scores by 0.38 standard deviations. In a Google blog post accompanying the paper, Autor and co-author Tanya Rodchenko described that result as an 11-percentile-point gain relative to the control group. The paper said the improvement came from a reduction in weak work rather than an increase in excellent work.

The more difficult test came at the end of the experiment. Lawyers had to mark up a hypothetical patent that, according to the blog post, contained numerous substantive and stylistic errors. The paper described the exercise as a task that patent lawyers “routinely perform unaided.” AI was not permitted.

Senior lawyers who had received AI access outperformed their control-group peers by 0.45 standard deviations, the paper found. Junior lawyers showed no average gain.

The study included several caveats. Only 91 of the 133 lawyers completed the AI-free test. The firms declined to administer a skills test at the beginning of the experiment, so the researchers inferred skill gains from random assignment rather than measuring skills before and after the trial.

The authors also flagged 15 of the 91 markups as possibly having been assisted by AI despite the ban. Such flags appeared about as often among control-group lawyers. After excluding those 15 markups, the paper said, the advantage for senior lawyers declined to 0.39 standard deviations and remained statistically significant at a looser threshold. The overall effect did not remain.

The “illusion of competence”

The performance of junior lawyers split across the distribution, according to the paper. There were significantly more poor scores, fewer mediocre scores and more good scores, but no improvement among the highest-scoring juniors. The authors wrote that AI “served as a springboard for some juniors and a cushion for others.”

The accompanying blog post described how junior lawyers used the system. They generally worked from the beginning of a document to the end, polishing introductory language before reaching the main claims. Some identified serious problems but left comments describing them instead of correcting them. Junior lawyers in the control group behaved similarly, a pattern the authors called “a baseline junior deficit” that three months of AI access did not eliminate.

The junior lawyers nevertheless reported a positive experience with the tool. Autor told Fortune that AI-assisted junior lawyers reported a surge in task satisfaction, the largest such effect in the study. They said that bypassing the “blank page problem” allowed them to move directly into the role of a reviewer, he said. That satisfaction stood alongside the absence of average gains in the unassisted test — a gap between how the work felt and what independent performance actually showed.

“Young professionals need to beware of the illusion of competence,” Autor 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 experienced lawyers gained

In follow-up interviews described in the blog post, experienced lawyers characterized AI as a “logic auditor,” not a finished product. The authors wrote that the tool weakened their attachment to existing prose and required them to articulate why and how they were making structural edits.

“Arguably, you can use AI as a ‘logic auditor’ only you already know the law well enough to spot when fluent text is legally flawed,” Autor told Fortune. He said that delegating baseline drafting allowed senior lawyers to focus on strategic scope — “the forest not the trees.”

During the unassisted test, the blog post said, treated senior lawyers skipped low-stakes prose, rebuilt claims from scratch, removed language that could narrow legal protection and connected many of their edits to legal doctrine.

Autor did not say junior lawyers should abandon AI. Instead, he told Fortune that novices should work more like senior lawyers by framing an argument without assistance first and then using AI as a critic to pressure-test the logic. He acknowledged that approach is difficult.

“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.

Protecting the apprenticeship pipeline

For employers, improved output could create an incentive to hire fewer junior workers. Autor called that approach 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 that firms distinguish between tasks that can be completed permanently with AI assistance and judgment-based tasks that require unassisted mastery. Companies could also combine AI deployments with regular unassisted skills checks, such as offline redlining exercises, and hold partners responsible for mentoring junior lawyers through the technology’s mistakes.

Autor emphasized that these are proposals rather than proven solutions. “It’s early days in this area of research and practice,” he said, adding that few approaches have been robustly validated.

The paper identifies additional limitations. The sample was small and came from firms performing ongoing patent work for a single sophisticated client. In addition, three months is a short period compared with the years required to develop patent expertise. The blog post said newer AI models and broader familiarity with the technology could produce different results if the experiment were conducted today.

The study does not establish that AI prevents junior lawyers from becoming experts over the course of a career.

Autor connected the issue to his earlier work on the China shock. Most technological advantage, he told Fortune, comes less from equipment or blueprints than from human know-how developed through slow and difficult 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.”

The story was originally featured on Fortune.com.