AI workplace learning is beginning to present an uncomfortable puzzle. In an experiment involving patent lawyers, access to an artificial intelligence assistant improved the quality of work produced. Yet when the lawyers later worked without the tool, the gain in independent professional judgement appeared mainly among experienced colleagues. Junior lawyers, on average, had not made a comparable leap, according to the MIT research team's working paper summary.
That is not proof that new workers should be barred from AI. It is evidence that getting a better document finished today and becoming better at the profession tomorrow are different achievements. Organisations adopting assistants need to consider both, especially when the routine assignments that develop expertise are the first to be automated.
AI workplace learning can hide a training problem
The study involved 133 practising patent lawyers at eleven US intellectual-property firms. Researchers randomly assigned access to a custom AI drafting assistant over roughly three months. The system helped with legal documents, while expert lawyers assessed the quality of work without knowing which experimental group had produced it.
Quality improved for those using AI. The improvement was visible during the trial, including at assessments after ten and ninety days. That is a genuine practical result for a demanding area of legal practice where small wording choices can have large consequences.
But the central question was what remained when the assistant was removed. The researchers asked participants to review and edit an existing patent application on their own. This exercise was intended to reveal judgement rather than the quality of machine-supported drafting. In the combined results the AI group performed better, but the advantage was concentrated in senior lawyers. The average junior result showed no such improvement.
The strongest performers were not necessarily the best learners
The data contained a more complicated pattern among less experienced participants. Instead of all junior lawyers benefiting or failing in the same way, their unaided scores spread towards both stronger and weaker performances. Researchers described a reduction in middling results, with more juniors at the good and poor ends of the distribution.
The finding matters because an average can conceal these differences. A workplace manager might see every AI-assisted draft looking competent and conclude that the entire team is progressing. The junior who is learning to interrogate suggestions could indeed be improving. Another junior may be accepting those suggestions without understanding the reasoning that a professional must eventually provide unaided.
The study cannot identify every reason for this split, and it does not establish that using AI caused permanent damage to anyone’s skill. It does suggest that managers need more than output quality to tell whether a new colleague is becoming independently capable.
Experienced workers may know which questions to ask
A senior lawyer has already spent years encountering rejected arguments, unusual cases and the consequences of poor wording. That background provides an internal standard against which an AI suggestion can be judged. If the software proposes a drafting change, the experienced lawyer may recognise both its value and its weakness.
A junior faces a different challenge. An answer can look authoritative before the learner has acquired the knowledge needed to challenge it. Copying a useful example is often part of learning, but copying without understanding becomes a problem when no one checks the reasoning behind it.
The authors’ discussion of the trial frames this as a difference between using AI as an aid to existing judgement and using it before that judgement has developed. The contrast should sound familiar to anyone who has learned a trade by doing small, repetitive jobs under supervision. Those tasks may appear inefficient precisely because their deeper value only becomes clear later.
Why entry-level tasks are worth protecting
Businesses have strong reasons to automate routine drafting, report preparation, sorting and initial research. Such work can be tedious, and experienced employees may want to spend their time on harder decisions. There is little sense in making a person repeat an administrative process solely to prove that they can suffer through it.
But some routine activity is also where apprentices notice exceptions, learn to check evidence and discover how errors are corrected. When an assistant supplies the polished result, a learner might miss the intermediate reasoning. The risk is not that a junior produces a good document too quickly. It is that nobody can tell what the junior contributed or learned.
LiveAIWire has looked at the quality of work with AI and the possibility that AI can improve rather than merely speed up work. That opportunity is real. The missing question is whether the improvements continue as a worker takes on more responsibility, with less supervision and fewer automated hints.
The practical response is to redesign the apprenticeship
A sensible training programme can let juniors use AI while preserving opportunities to make and explain their own decisions. A trainee could draft an argument first, compare it with a suggested alternative and explain why one is stronger. A supervisor can question the assumptions instead of grading only the finished document.
Another useful method is occasional unaided assessment. If a person cannot solve a problem without the tool that usually solves it for them, that is valuable information about the training programme. It is not a reason to humiliate the learner; it identifies what they need to practise. Assessments should reflect genuine job demands and be proportionate to the stakes.
Teams can also review AI mistakes together. Understanding why a fluent paragraph is misleading may provide a stronger lesson than repeatedly accepting accurate suggestions. LiveAIWire’s AI closing workplace performance gaps reporting on workplace performance provides useful context for this tension between immediate output and underlying ability.
One legal experiment does not settle every profession
The trial involved US patent lawyers and a customised legal drafting assistant, not every office worker using a general-purpose chatbot. The participants were experienced professionals at different career stages, and the outcome measure focused on a particular kind of judgement. It would be unsafe to claim that the same pattern must apply to schoolchildren, accountants, programmers or engineers.
The study is also a working paper rather than a peer-reviewed conclusion accepted across an entire field. Its design has strengths, including random assignment, a period of ordinary use and blind expert grading, but replication across occupations and products remains important. The authors disclose connections to Google, including support and staff involvement. That is a reason for transparency and scrutiny, not automatic grounds for rejecting the evidence.
Organisations should treat the result as a warning to measure learning properly. The headline productivity number is only one part of what determines whether a workplace remains competent in future.
What supervisors should measure instead of speed alone
A useful measure of workplace learning would include the quality of reasoning, the ability to notice an exception and the worker’s confidence in explaining a decision without the assistant. Supervisors might ask a junior to identify an AI error, describe the evidence that would change an answer and explain the consequences of choosing one option over another.
These checks need not recreate every old-fashioned form of training. Repetitive clerical labour has costs of its own, and an employee who is constantly deprived of helpful tools may learn less, not more. The point is to deliberately preserve the experiences that build judgement and remove the tasks that provide little educational value.
Managers should also ask whether a worker’s improvement persists. A team that performs well only while a particular subscription or tool remains available could be more fragile than its productivity figures suggest. Evidence of independent competence is an investment in resilience as much as an assessment of education.
A better assistant might sometimes refuse to finish
One possible design implication is an AI assistant that behaves differently for a learner and an expert. The expert may benefit from a concise recommendation and a quick route to a finished draft. The trainee may gain more if the system asks for an initial judgement, offers a hint, or explains why an apparently attractive answer is weak.
That would require a careful balance. Unnecessary obstacles can frustrate workers, and no single learning style suits every person. Nonetheless, treating training as a legitimate product goal could improve the long-term value of workplace AI. Firms might assess products partly on whether staff become better at difficult decisions rather than only on how many documents are completed.
The experiment is useful because it exposes the risk of confusing a polished deliverable with an increase in understanding. It gives researchers, employers and software companies a question they can actually test.
A workforce cannot be made only of finished experts
It is easy to imagine an employer using AI to raise the quality of every outgoing document. It is harder to imagine the same employer thriving if nobody learns to handle the rare case the software gets wrong. Companies need a route from beginner to expert, not merely a set of tools that allows today’s experts to produce more.
That concern connects with LiveAIWire’s earlier coverage of the value of senior workers and the shifting value employers place on experienced staff. A company that rewards experience while eliminating all the work through which experience is earned could eventually create its own shortage.
The useful question for the next wave of automation is therefore not simply how many hours AI saves. It is which moments of real judgement, correction and responsibility remain for the humans who will have to make tomorrow’s 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.