There is a growing body of research on AI in education, but strikingly often it remains at the same level: does it work better with or without AI? That, in itself, is an interesting question, but it misses something fundamental. It says little about what actually changes in teachers’ work.
A recent preprint by Kramer and colleagues, which I came across via co-author Inge Molenaar, tries to approach this differently. Instead of looking at effects, they look at tasks. What happens to the concrete actions of teachers when AI is introduced? And what does that mean for the knowledge and skills required?
To do this, they combine two analytical frameworks. On the one hand, they use a task analysis, breaking teaching down into goals, subtasks, and actions. On the other hand, they draw on an analysis of the underlying skills and knowledge required at each step. The aim is therefore not to prove whether AI works or not, but to make visible how the work itself shifts.
This leads to two interesting cases.
In the first case, the focus is on an AI tool that helps plan differentiated mathematics lessons in primary education. Without AI, this is a fairly linear process: the teacher analyses the class, selects a goal and creates a plan. With AI, that overall goal remains the same, but the structure changes. The tool generates suggestions based on student data, and the teacher must interpret, evaluate, and, if necessary, adapt them. The process becomes less of a straight line and more of a series of decision points.
In the second case, involving generative AI for feedback planning in secondary education, the work shifts one step further. Here, the emphasis is not on evaluating AI output, but on feeding it. Teachers need to make their own knowledge explicit in prompts and then collaborate with the AI in multiple rounds to arrive at usable feedback strategies. The work becomes more explicit, more iterative, and also more dependent on how well one can articulate one’s own thinking.
The authors describe this as two forms of complementarity. In the first case, AI supports decision-making; in the second, the teacher’s expertise serves as input to guide the AI. In both cases, the conclusion is remarkably consistent: AI does not replace teachers’ thinking, but reorganises it. Certain analytical tasks are shifted to the system, while interpretation, context, and, above all, responsibility remain with the teacher.
For me, that is the paper’s strong point. It makes concrete what often remains abstract. Instead of general claims about “AI as a partner”, the paper shows where and how that collaboration takes shape.
At the same time, there are also a few tensions, in my view. Bear with me…
By focusing so strongly on tasks, actions and decisions, teaching is almost inevitably approached as something that can be neatly decomposed. That works well for analysis, but the risk is that an important part of practice falls out of view. Teaching is not just a sequence of actions. It is also relational, context-dependent and often messy. Things like timing, classroom dynamics or sensing what is needed at a particular moment are difficult to capture in a task model.
Related to this, the paper leans strongly towards a primarily cognitive perspective. The emphasis is on interpreting, deciding and evaluating. What becomes less visible is that much of teachers’ expertise is implicit. Not everything a good teacher does can simply be made explicit, let alone turned into a prompt. For some, that may be the problem; for others, it is part of the profession’s beauty.
There is also a fairly optimistic view of practice underlying the analysis. The described forms of complementarity assume teachers who have the time to evaluate AI output, sufficient knowledge to understand it, and who remain critical. In reality, we know these conditions are far from always present. In such cases, complementarity can quickly shift towards automation or simple convenience.
And perhaps more fundamentally: by focusing on who performs which task, the discussion almost imperceptibly shifts away from the question of what education is actually about. It becomes about efficiency and task allocation, rather than goals, values, and what we consider important for students to learn. It may be that all this could create more space for what truly matters, but that is (still) uncertain.
Don’t get me wrong. I do not want to attack or dismiss the solid work of the researchers, but rather engage with it and think further. This does not make the paper any less interesting. On the contrary. It does something we need: it brings structure to a debate that often remains too vague. But as with any model, it is a lens. And like any lens, it makes some things sharper, and others less so.