As AI increasingly becomes part of professional practice, we may need to rethink not only how people use it, but how people learn to become, and remain, capable practitioners.
In a recent post I asked what better work with AI might look like. That question leads to another: what would better education, training and professional development for AI-supported work look like?
Much of the current discussion about AI and education starts with cheating. That is understandable. Schools and universities need to know whether submitted work represents a student’s own effort, and qualifications need to mean something. But there is a risk that we hold the existing educational model constant. We continue to set much the same essay, assignment or task, then focus on how to prevent students using AI to complete it.
That may address an immediate problem without getting to the larger one. If AI is going to be part of the work people eventually do, should we also reconsider the learning goals, activities and assessments themselves?
This is a different question from how AI might improve existing education. AI may provide better tutoring, faster feedback, help with drafting, or easier access to information. All of those may be useful. But they largely ask how AI can help us do familiar forms of education and training better.
A more fundamental question is what people need to learn differently when AI becomes part of the practice for which they are being prepared.
The distinction matters well beyond schools and universities. AI is becoming part of research, policy, medicine, engineering, evaluation, law, teaching and many other kinds of professional work. It is increasingly used to search, draft, summarise, compare, analyse and suggest possible interpretations. Professional development is changing too, because experienced practitioners are now learning while working alongside systems that can take on parts of activities through which they once developed and maintained their expertise.
There is an important difference between being able to use AI and being capable of using it well.
A person can learn how to prompt a system, ask for alternatives, request a summary or generate a first draft quite quickly. But using AI well depends on much more than operating the tool or checking its outputs. It depends on understanding the work itself: what matters, what good work looks like, what can reasonably be inferred, where context matters, and when further inquiry, judgement or discussion is needed. That wider understanding is what allows someone to decide where AI can help, where it may mislead, and where it should remain only one part of the process.
Those capabilities do not appear simply because AI is available. They have to be developed somewhere. This connects with work by Bearman and colleagues on evaluative judgement, including how people learn to judge both generative AI outputs and the processes used to produce them.
That raises questions about what people still need to understand, practise and experience for themselves. It also raises questions about activities that may look productive on the surface but are doing developmental work at the same time.
When doing the work is also part of the learning
We write a report because a report is needed. We analyse a set of data because we need the result. We discuss a difficult case because a decision has to be made. But these activities often do two things at once. They produce an output, and they develop the people doing the work.
Writing can help us discover what we think. Comparing several cases can teach us which differences matter. Struggling with an analysis can reveal where our understanding is weak. Discussing an interpretation with colleagues can expose assumptions that would otherwise remain invisible. Supervising someone through a difficult piece of work can develop the judgement of both people involved.
What makes these activities developmental, though, is not always the task itself. Feedback, having to explain a judgement, encountering disagreement, or later discovering whether we were right can all matter. The question is therefore not simply which activities to preserve, but which conditions help people learn through them.
Suppose a multi-site evaluation has generated findings from a number of local projects. At some point those findings need to be brought together into a smaller number of wider conclusions. AI can help organise material, compare reports, identify recurring themes and suggest possible patterns.
That could save considerable time. It might also improve the process by making it easier to work across a large body of material.
But much of the judgement lies in making sense of that synthesis. Are two findings really describing the same thing? Is a difference between sites incidental, or does it change the meaning of the finding? Can something reasonably be generalised, or is it too dependent on context? Does an apparent pattern reflect the evidence, or simply the way the material has been framed?
As I discussed in an earlier post on the authority of the first synthesis, an early AI-generated interpretation can also quickly shape what is discussed next. Those questions therefore involve more than checking whether the synthesis is accurate. Some of the judgement is developed and exercised through the process of working them out together.
A group may begin with different interpretations and gradually come to understand why they differ. Someone may notice that a conclusion that seems obvious from one perspective looks quite different from another. A discussion may reveal an assumption embedded in the original questions. The learning is not separate from the work. It happens through doing the work, and sometimes through having to make sense of it with others.
If AI provides an early, plausible synthesis, that may be useful. But it may also shape the discussion that follows. Instead of asking openly what the material might mean, people may find themselves reacting to categories and patterns that have already been proposed.
The issue is therefore not whether we should use AI for synthesis. It is whether we understand what else was happening in the activity that AI is now helping us perform.
That question matters for curriculum design, training and workplace learning. If an activity once produced both an output and a learning experience, we need to know whether handing more of it to AI changes the learning as well as the workload. If it does, we then need to ask whether that learning still matters and, if it does, where it will happen instead.
Developing judgement, and keeping it
This is not only a question about novices.
Experienced practitioners may be particularly good at working with AI because they have years of practice against which to judge what it produces. They have seen enough cases, mistakes, exceptions and awkward situations to recognise when something does not quite fit.
There is a transition problem here. Many people who are good at judging AI-supported work developed that judgement through years of doing work that AI may now partly take over. Where will the next generation get the experience on which that judgement depends?
Judgement is not simply accumulated and then stored away. It also has to be exercised, tested and renewed, while also being kept current as the practice itself changes.
If AI increasingly takes over parts of the activities through which practitioners encounter difficult cases, compare alternatives, work through uncertainty or explain their reasoning to others, what happens to that judgement over time?
This suggests a distinction between developing judgement and maintaining it. A novice needs opportunities to build it. An experienced practitioner needs opportunities to keep using it.
The answer cannot simply be to insist that people continue doing everything manually. Some activities may no longer deserve the time they once consumed. AI may remove routine work that was never especially developmental in the first place. It may also create new learning opportunities, for example by offering rapid feedback, generating alternative explanations or making it easier to explore several approaches to a problem.
Working with AI can itself be developmental if people have to form their own view, compare it with what AI produces, and explain where and why they differ. The more useful question is which experiences matter for capability, and why.
Where does judgement move?
That also changes how we think about keeping a human in the loop. The phrase can make it sound as though judgement remains in the same place and a person simply checks the machine’s work before it moves on.
In practice, judgement may move.
It may move to the person who frames the initial question. It may be shaped partly by assumptions embedded in the system, and by the categories or patterns the system makes readily available. It may move downstream to whoever decides whether an output is acceptable. Or a judgement that once emerged through discussion among several people may shift into an individual interaction between one person and an AI system.
In the synthesis example, this could mean judgement shifting from a group working through interpretations together to one person framing an AI request and deciding whether its resulting synthesis is good enough.
That change matters for learning as much as it does for accountability. It may also matter for collective sense-making. A process that still contains human judgement is not necessarily the same process if that judgement is now exercised individually rather than worked through with others. A recent study by Hussain and Sveen similarly found that teams evaluating AI outputs developed their judgements through discussion and shared interpretation, rather than individual checking alone.
For schools and universities, this suggests looking beyond whether students used AI. We also need to ask what a task was intended to develop, whether that learning still matters, and whether the same task remains the best way to develop it.
For professional preparation, the question is somewhat different. Where should people learn to work with AI, and where do they still need direct experience of the underlying practice because important capabilities depend on that experience?
For continuing professional development, the focus shifts again. The question may be less about courses than about the work itself. Where will experienced practitioners continue to encounter uncertainty, compare interpretations, discuss difficult cases, test their reasoning and learn from each other as AI takes on more of the routine work?
There is unlikely to be one model for this across different fields. What a surgeon needs to learn through practice is different from what a policy analyst, teacher, engineer or evaluator needs to learn. Nor will every activity that once had educational value need to be preserved.
But a small set of questions may help us think about what needs to change.
- What capabilities do we want people to develop and maintain?
- What do they still need to practise or experience for themselves?
- Which parts of the work can AI usefully take over?
- If an important learning experience changes or disappears, where will that learning happen instead?
- And where does judgement move when AI becomes part of the process?
If AI is changing professional practice, then improving existing education with AI is only part of the task. We may also need to work backwards from changing practice and ask what forms of education, training and professional learning will help people become, and remain, capable practitioners within it.
For more on this wider theme, see my working well with generative AI collection, which brings together related essays and resources on AI-supported professional practice.
[* Image by tippapatt / Adobe Stock]
