The greatest risk may not be that today’s graduates lose their jobs but that tomorrow’s senior professionals never properly learn how to do them.
I have some reservations about newly graduated psychologists using AI to write their clinical notes.
For an experienced psychologist who has written thousands of notes, having AI generate a decent first draft in seconds is an obvious efficiency. But I’m much less enthusiastic about someone using it from their first day on the job.
Writing clinical notes is tedious, but it is also part of learning to think like a psychologist. After a session, a young clinician has to decide what mattered, what was observed rather than reported, whether there is a risk issue and what needs following up. They have to sort the important from the incidental.
The note is one product of that exercise. The other product is the clinician.
The same dilemma is now playing out across graduate work more broadly.
Earlier this year, MinterEllison revealed it was reducing its graduate intake from 104 positions to 72, with efficiencies created by AI among the reasons cited. Some of the routine work traditionally handed to junior lawyers can now be done far more quickly with technology.
You can hardly blame a law firm for taking advantage of that. Nor should we preserve pointless work simply because previous generations had to endure it. If a task that once occupied a graduate for half a day can reliably be done in minutes, making them do it the old way indefinitely would be absurd.
Still, I wonder whether we have been too quick to assume that the only value of junior work was its output.
A junior lawyer reading a mountain of cases isn’t just producing research for somebody more senior. Along the way, they learn how judgments are constructed, which facts matter, and what changes an argument. A lot of professional training has always happened this way, without anybody particularly designing it.
The dilemma is complicated by evidence that AI appears to be particularly useful for inexperienced workers.
A study of more than 5000 customer-service workers found generative AI increased productivity by 15 per cent overall, with the biggest gains among less experienced and lower-skilled workers. Remarkably, AI-assisted workers with just two months on the job performed as well as unaided workers with more than six months’ experience.
For an employer, that is an extraordinary proposition. You can take someone with limited experience and rapidly improve the quality and speed of their work.
There is a catch, though. Better performance today doesn’t necessarily tell us what that worker is learning for tomorrow.
Medical researchers have recently used the term “never-skilling” to describe one possible problem. Unlike deskilling, where technology erodes an ability someone already has, never-skilling occurs when a person outsources a task before they have properly developed the skill themselves.
This is where the distinction between performance and expertise becomes important. AI allows a novice to produce work that can look considerably more sophisticated than their own level of expertise.
An experienced professional has years of knowledge against which to judge what AI produces. They are more likely to notice what has been missed, when something doesn’t quite fit, or when an apparently convincing answer is simply wrong. A graduate may have far less ability to make that distinction.
This doesn’t make AI bad for learning. There is evidence from medical education, for example, that supervised and thoughtful use of AI can improve learning and critical thinking. The issue is how and when it is used.
We also shouldn’t get ahead of ourselves on jobs. Australia has not experienced the mass AI-driven unemployment that has been predicted since ChatGPT arrived. Jobs and Skills Australia has found only limited displacement so far, with AI instead changing roles and creating a need for upskilling and redeployment.
But there are signs that younger white-collar workers deserve particular attention. Recent Stanford research using payroll data from millions of US workers found employment among 22-to-25-year-olds in highly AI-exposed occupations was about 19 per cent below where it would have been had it kept pace with their less-exposed peers. Experienced workers showed no comparable gap.
Perhaps that isn’t surprising.
Much of the value of an experienced professional lies in things that are difficult to put into a prompt. They have seen a version of the problem before. They notice when something feels off. They know when they need more information and, importantly, when not to trust the obvious answer.
We usually call that judgment, although describing how somebody acquires it is much harder.
For decades, professional firms didn’t really have to. Graduates joined at the bottom, did fairly mundane work, watched more experienced people, got things wrong and gradually took on more responsibility. Some eventually became the senior lawyers, partners, consultants, doctors, accountants and executives whose judgment organisations now value so highly.
AI makes that pathway less straightforward.
Hybrid work has already reduced some of the incidental observation through which graduates learned from senior colleagues. AI may now remove some of the repetitive doing as well.
None of this is an argument for keeping graduates technologically handicapped for the first five years of their careers. They need to learn how to use AI, probably much better than most of their bosses currently do.
But employers may have to think more seriously about what disappears along with the grunt work. Which supposedly menial tasks were also training people, and how will that capability now be built?
For businesses, there is an obvious incentive to do this. Saving time and reducing the number of graduates required looks good on this year’s numbers. The harder question is who becomes your senior talent 10 or 15 years from now.
The greatest risk of AI may not be that today’s graduates lose their jobs. It may be that tomorrow’s senior professionals never properly learn how to do them.
