David Autor and six co-authors ran the experiment that almost nobody running AI pilots bothers with. They gave 133 practicing patent lawyers at eleven American intellectual property firms a custom AI drafting assistant, left it in place for three months, and then took it away and tested them without it.1

That last step is the whole study. Nearly everything published about AI at work measures what people produce while the tool is switched on, which answers a question about the tool. Removing it answers a question about the person.

Two details worth noting before the results. The trial was pre-registered before any data came in, so the hypotheses were fixed in advance rather than found afterwards. And every piece of work was graded by expert patent attorneys who did not know which lawyers had been given the assistant.1

With the tool switched on, the finding is the familiar one

Drafting quality went up. At ten days, work from lawyers with AI access scored 0.34 standard deviations above the control group. At ninety days it was 0.38.1 The gains were larger among the junior lawyers, which is both what earlier studies in other white-collar fields have found and what most people would predict. The tool lifts the least experienced the most.1

Then they took it away

At the end of three months, every participant redlined a real patent application with no assistance of any kind. Redlining is core patent practice and it runs on judgment rather than on production.1

Taken as a whole group, the lawyers who had spent three months with AI beat the control group by 0.32 standard deviations.1 Stop reading there and the story is that AI practice builds durable skill.

Split the same result by seniority and it comes apart. The entire advantage sat with the senior lawyers, who scored 0.45 standard deviations above their controls. The junior lawyers showed no average gain at all.1

The authors summarize it in a sentence I have not been able to put down since I read it.

The largest gains from AI thus accrued to the lawyers who retained the least.

The juniors did not simply stay flat

An average of zero hides the more interesting thing that happened. Junior scores bifurcated. There were sharply fewer mediocre results, offset by more poor ones and more good ones.1 The group did not hold still, it pulled apart.

The paper does not settle why, and I am reading past the evidence here, but the shape is recognizable. Two people can spend three months with the same assistant and use it in opposite ways. One argues with the draft, checks it, works out where it is wrong and why. The other accepts it and moves on. While the tool is running, those two produce almost identical output and look equally productive on any dashboard you care to build. Take the tool away and they are not the same lawyer at all.

Who paid for it

This matters enough to state plainly. Google funded the direct costs of the experiment, six of the seven authors are Google employees, and Autor's support includes a Google fellowship program.1 That is a real interest and you should weigh it.

It happens to cut against the funder here. The headline result is uncomfortable for anyone selling AI as a way to get more out of junior staff, and a study bent toward its sponsor would have landed somewhere else. Worth knowing all the same.

What a leader does with this on Monday

The standard deployment plan is to hand AI to the people whose work is most routine and then count the throughput. On this evidence that plan buys real output now and nothing durable from the people you were relying on to become senior.

The authors' own conclusion is carefully worded: foundational expertise may be a prerequisite for extracting durable skill from AI-assisted practice.1 The word doing the work is prerequisite. If foundations have to be in place first, then the order of a rollout matters more than its speed, and most rollouts are sequenced the other way round.

I wrote in June about companies handing junior work to AI and thinning out the bench that produces senior people. That argument came from hiring and payroll data. This is the same worry arriving from the other direction, with a control group. That piece asked whether you hire juniors at all. This one asks what happens to the ones already sitting on your team.

What would make me doubt it

133 people in one profession over three months is not a large study, and it is a working paper that has not been peer reviewed. Patent drafting is unusually structured work, so the result may not travel to messier jobs. Ninety days is a short window for a claim about expertise, which is normally built over years, and a three-month reading could easily miss a slower catch-up.

What survives all of that is the design. Almost nobody switches the tool off and measures what is left, and the one team that did found the answer depends on who was holding it.

So here is the question worth carrying into your own organization. Think of the person on your team who has leaned on AI hardest this year. If it disappeared on Monday, would they be better at the actual work than they were in January, or only faster while it was running? Most organizations have no way to answer that, because nothing they measure ever turns it off.