Does AI make junior staff better, or only make their work look better?

Better at the work they hand in, on the evidence so far. Better at the job has not been shown. In the one trial that has tested both, 133 patent lawyers were followed for three months. With an AI drafting assistant everybody’s work improved, and the juniors’ improved most. Then all of them marked up a patent application without it. The lawyers who had used the tool did better than those who had not, and the whole of that advantage sat with lawyers of seven or more years’ experience. Juniors showed no average gain. Google paid for the trial and it has not been peer reviewed. It is still the first to separate two things an employer usually counts as one: what a junior produces, and what a junior can do.

What the trial did#

The working paper, by David Autor and six co-authors, was issued by NBER in September 2026 and written up by Google Research on 7 October. It was pre-registered. Two-thirds of the lawyers at each of eleven US intellectual property firms were given an AI patent drafting assistant that Google had not yet released. The rest were given some training in using AI and no tool until the three months were up. Expert patent attorneys scored everyone’s work without knowing which group it came from.

While the tool was in use the result was the familiar one. Drafting quality rose by 0.34 standard deviations at ten days and 0.38 at ninety, “with larger gains among junior lawyers”, who were also faster: on the ten-day task, about 18 minutes quicker than a control average of 124. The same pattern, with the least experienced gaining most, was found in customer support in 2023.

What happened when the tool was taken away#

At three months every lawyer redlined an existing patent application with no AI, a task the paper calls “a core task of patent practice requiring expert judgment”. Those who had spent the three months with the tool beat the control group by 0.32 standard deviations, “but this advantage was concentrated entirely among senior lawyers”, at 0.45. For juniors, defined as under seven years in practice, the abstract reports “no average gain; their scores instead bifurcated, with sharply fewer mediocre scores offset by more poor and more good ones.” Its last two sentences: “The largest gains from AI thus accrued to the lawyers who retained the least. Foundational expertise may be a prerequisite for extracting durable skill from AI-assisted practice.”

The interviews suggest why. Seniors described treating the tool’s output as a “logic auditor”, something that made them say why an edit was right. Juniors, the authors write, often spotted a serious flaw and left a comment describing it instead of making the fix. That habit was just as common among juniors who never had the tool, so the tool did not cause it. Three months of having rewrites done for them did not cure it either.

Why it matters to whoever hires and promotes#

An employer sees output. In this trial the output of juniors with the tool rose furthest, and nothing in the output showed that their unaided judgement had stood still. This research calls that gap synthetic seniority: work that looks senior while the judgement underneath has not been built. Employers appear to sense it. In a survey experiment on about 1,750 US professionals with hiring responsibilities, telling them that students use AI raised their concern about evaluating applicants, and of the kinds of evidence offered to reassure them “only in-person tests were selected by both treatment groups.” Lyndon Nicholson, founder of Future Group, described in Consultancy.uk on 2 October a manager whose polished case stopped at one question from the leadership team, “What happens if this doesn’t work?” His line on it: “The deck had done the thinking’s job so well that he’d never needed to.” That is one anecdote in an opinion piece. The trial is the measurement.

The evidence that points the other way#

Two studies published the same week show juniors gaining. A field experiment with 673 final-year IT apprentices at twelve German vocational schools found that with AI “productivity increases by 26–31 percentage points on tasks beyond apprentices’ formal training”, and the authors add: “These productivity gains do not come at the cost of reduced immediate task comprehension.” Comprehension was tested at once and not months later. In England, a randomised trial of an AI lesson planner with 464 primary teachers cut planning time by about a quarter, with lessons found “not to be of lower quality”, and new teachers and those less confident in their subject were the likeliest to see their workload fall. Ben Styles of NFER, which ran it, put the open question in one sentence: “Substituting for experience could really help or, alternatively, it could risk poorer-quality lessons.” Neither study measured what the junior could do afterwards without the tool.

What seems to separate the juniors who gain#

The patent juniors split, and two smaller studies suggest what the split may follow. In two preregistered experiments in which people ran a simulated factory with an AI adviser, those who more often altered its recommendations went on to do better unaided in the first study and knew more in the second. A study of 110 novices on a clinical text task found “systematic drift toward over-reliance in the presence of explanations”, with more selective use among those who reported understanding the task. Set beside the seniors’ “logic auditor”, the common thread is arguing with the output instead of accepting it. That reading is an interpretation. Both smaller results are associations, and neither has been peer reviewed.

What an employer controls#

Three decisions, none of which waits for more research. First, measure the two things separately. The trial’s authors say that understanding the effect of AI on professional skill “requires tests that separate the effect of using tools from the effect of unassisted user performance”. A firm can do what they did: at intervals, a real task with no tool, marked by someone who does not know who used what. The method is on assessing capability rather than output. Second, decide which early tasks juniors still do by hand, and write that down as part of custody. If foundational skill is the precondition for learning from the tool, it has to be built somewhere. Third, say who supervises. The Solicitors Regulation Authority opened a consultation on 8 October on updating its statement of solicitor competence in areas that include “the use of technology and AI” and supervision, open until 3 December.

These sit inside the larger decision an organisation controls: which decisions a machine may make, who can stop each one, what people must remain able to do, and how anyone would know if it went wrong. That is Rules Before Tools. The longer account of where junior work goes is on the missing rungs, and the cost of leaving it unmeasured is on capability debt.

What this does not show#

It does not show that AI harms juniors. No group in the patent trial did worse on average, and the juniors’ scores spread in both directions, with more good ones as well as more poor ones. It is one trial of 133 lawyers at top-tier firms over three months, and the authors say the tools have moved on since. Google paid the direct costs, five of the seven authors are Google employees and a sixth is a paid contractor, the tool was Google’s and the firms do business with Google. The paper is a working paper, the sizes of the junior and senior groups are not given in the account read here, and nobody has followed a cohort that began its career with the tool, the gap recorded on whether deskilling is real. Whether the result holds for accountants, engineers or teachers is unknown. The apprentice and teacher studies show real gains for juniors and did not test them later without the tool, so they neither confirm nor contradict it.