A three-month experiment with 133 practising US patent lawyers produced a result that complicates the usual debate about AI for patent lawyers and professional skill. AI helped patent lawyers produce better work while they were using it, with the largest immediate gains among junior lawyers. But when the AI was taken away, durable improvement appeared among senior lawyers, not juniors.
The pre-registered randomised controlled trial involved lawyers at 11 intellectual-property firms and a custom AI drafting assistant. Their work was scored by expert patent attorneys who were blinded to treatment status. The question was not simply whether AI improved output. The researchers wanted to know whether practising with the tool built expertise that survived when the tool was no longer available.
AI for patent lawyers improved work while the assistant was present
On benchmark patent-drafting tasks, lawyers with access to the AI produced higher-quality work after 10 days and again after 90 days. The paper reports improvements of 0.34 standard deviations at the earlier test and 0.38 standard deviations at the later one. Junior lawyers benefited more while using the assistant.
That part of the result fits a familiar pattern. Less experienced professionals often have more room for a tool to help with structure, completeness or routine drafting. An assistant can expose them to stronger examples and reduce the cost of producing a competent first version.
The more surprising part arrived after three months. All participants were asked to redline an existing patent application without AI, a task designed to require professional judgement. The treated group performed better than the control group overall, but that advantage was concentrated among senior lawyers.
The MIT Stone Center summary reports that treated senior lawyers scored 0.45 standard deviations above senior controls on the no-AI task. Junior lawyers showed no average improvement. Their results became more spread out, with fewer middling scores but more scores at both the low and high ends.
The people who gained most from AI retained the least
That is the study’s central paradox. Junior lawyers received the largest boost from the tool during AI-assisted work, yet the senior lawyers showed the clearest evidence of retained improvement after the assistant was removed. The researchers suggest that foundational expertise may help professionals convert AI-assisted practice into durable skill.
One possible interpretation is that experts know what to notice. A senior lawyer can compare an AI suggestion with a rich internal model of good patent work, recognise why one formulation is stronger and integrate that lesson into future judgement. A junior may successfully use the suggestion without yet possessing enough background knowledge to extract the same general principle.
That explanation is plausible, but it remains an interpretation rather than a direct measurement of what happened inside each participant’s learning process. The experiment shows a difference in later performance. It does not prove a single cognitive mechanism behind the difference.
LiveAIWire has previously reported on evidence that AI assistance can weaken some human skills. The patent-lawyer trial is more nuanced. It does not show universal deskilling. It suggests the same AI-supported practice can produce different learning effects depending on the expertise a person brings to it.
A tool can improve the product without improving the professional
This distinction matters far beyond law. Employers usually evaluate AI by looking at output: was the report better, was the code completed faster, did the document need fewer corrections? Those are sensible measures of productivity. They do not reveal whether the employee is becoming more capable of doing the underlying work independently.
In many professions, that may not matter every time. Calculators improve arithmetic output even if a user becomes less practised at mental calculation. But professional judgement is different when people are expected to catch unusual errors, supervise automated systems or perform under conditions where the tool is unavailable.
Patent drafting is a useful test because it combines structured language with specialist judgement. The AI can help produce text, but a lawyer still needs to understand the claims, identify weaknesses and assess whether wording protects an invention appropriately. The later redlining task was designed to probe that judgement without the assistant.
LiveAIWire has also covered AI tools used to help people understand legal judgments. Both cases show why legal AI cannot be reduced to text generation. The valuable part often lies in what a trained person recognises about the text, not in the surface fluency of the document itself.
Junior training may need to change if AI does more of the first draft
The finding raises a difficult question for firms: how do people become senior experts if AI makes junior work easier before the underlying skill has fully developed? Traditional professional training often relies on repeated exposure to imperfect drafts, corrections and feedback. If a model removes too much of that struggle, the path to expertise could change.
The answer does not have to be banning AI for junior staff. The trial itself found clear productivity and quality gains while the assistant was present. Instead, organisations may need to design AI use as a learning process. That could mean asking juniors to critique generated drafts, explain why changes were made or complete selected tasks without assistance before comparing their work with the model.
Senior lawyers may also benefit from deliberate reflection rather than passive acceptance. Their retained gains suggest experience can turn AI interaction into learning, but the study does not establish that every expert automatically gets better. A tool can still encourage shallow review if speed is rewarded more than understanding.
The same challenge appears in other knowledge jobs. LiveAIWire has examined how AI adoption can change the balance between junior and senior hiring. If firms simultaneously hire fewer juniors and automate more of the work that once trained them, they may eventually face a thinner pipeline of experienced people able to supervise the systems.
The experiment is strong evidence, but it is still one profession
The design is stronger than a simple survey because participants were randomly assigned and their work was scored blind. It also ran for three months, long enough to ask whether repeated use changed later unaided performance. Those features make the result unusually useful in the debate over AI and expertise.
There are still limits. The sample was 133 patent lawyers in 11 US firms, using a custom drafting assistant. Patent practice has its own routines, incentives and knowledge structure. The results cannot be assumed to apply unchanged to accountants, doctors, engineers, teachers or software developers.
The study also measures one form of retained judgement after a relatively short period. Expertise develops over years. Longer studies could show different effects, especially as workers adapt their habits or models become more capable.
The real question is what AI teaches while it helps
The most useful way to read the experiment is not as evidence that juniors should avoid AI or seniors are safe from deskilling. It is evidence that better output and better learning are separate outcomes. A system can deliver the first without guaranteeing the second.
That distinction should change how organisations evaluate professional AI. Productivity dashboards can measure time saved and documents completed. Training systems need additional measures: whether people can identify mistakes, explain decisions and perform critical tasks without the model.
AI assistance is likely to become normal in legal work because the immediate gains are too useful to ignore. The harder design problem is making sure the people using those tools are still accumulating the judgement they will need later. The patent-lawyer experiment suggests experience can turn AI into a teacher, but only if the human already has enough knowledge to learn from what the system is doing.
The trial also suggests that firms should separate “assistance” from “apprenticeship” when designing work. If a junior lawyer always sees the model’s polished answer before forming an independent view, the AI may become a shortcut around the reasoning that builds expertise. If the same lawyer first makes a judgement, then compares it with the assistant and receives feedback from a senior, the tool may support learning rather than merely replace practice.
That possibility was not directly tested here, so it should be treated as a design hypothesis. It is nevertheless testable. Employers can compare training formats, measure unaided performance over time and identify which kinds of AI use leave people better able to explain and defend their decisions.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.