Is AI doing the thinking your pupils need to do?

A pupil asks an AI tool to explain photosynthesis and receives a clear, accurate paragraph in seconds. The explanation now exists. The question for school leaders is whether any learning has happened.

A new expert consensus report from the University of Minnesota, Ways Generative AI Can Support and Threaten Learning in K-20 US Education, suggests the answer depends almost entirely on who does the thinking. AI has a potentially useful role in practice, recall and consolidation. It becomes a risk when it allows pupils to avoid the effort that learning requires.

What I particularly liked is the report's method. Thirty experts examined 65 learning processes and judged whether AI could enhance each one, diminish it, or do both. That gives leaders a better starting point than most edtech debates: specify the learning process before judging the technology. It is worth being clear that this is structured expert judgement, not a trial proving the impact of a particular product.

What the experts found

Practice and consolidation offer the clearest opportunities. Self-testing, feedback, worked examples, questioning and deliberate practice were all identified as areas where AI could help strengthen knowledge pupils have begun to acquire.

Avoiding effort can mean avoiding learning. If AI supplies the explanation, retrieves the answer or completes the reasoning, pupils may miss the mental work that builds knowledge. Prompting a pupil to recall something is very different from recalling it for them (duh).

Higher-order thinking needs particular caution. The experts identified risks to critical thinking, evaluation and inquiry. Their judgement does not support assuming that current general-purpose AI reliably develops these capabilities (with currently available tools).

Younger learners and learning relationships need protection. There was greater concern about primary-aged pupils, alongside risks to self-regulation, collaboration and social connection.

The same process can benefit or suffer, depending on use. Motivation, for example, attracted consensus on both potential enhancement and potential harm.

These are expert judgements rather than measured effects, and stronger evidence of actual learning impact is still needed. Even so, they point in a direction leaders in England will recognise. In January 2026 the DfE updated its Generative AI: product safety standards with new standards on cognitive development, responding to the risk of cognitive offloading. Products are now expected to report to teachers on how often pupils ask the tool to do their thinking for them. The US experts and the DfE have reached the same concern from different directions.

Two Year 4 classes

Imagine two Year 4 classes studying the water cycle, both using AI.

In the first, pupils are taught the content. They then use a tool that asks them questions. They answer in their own words before seeing any feedback. When they are stuck, they get a hint rather than the answer. A week later, the tool brings the same ideas back.

In the second, pupils ask the AI to explain evaporation and condensation, then write up the explanations in their books. Their work is neater, more accurate and finished sooner.

On a learning walk, the second class might look more impressive. But close the tablets a fortnight later and ask both classes why puddles disappear on a sunny day. I know which class I would expect to explain it.

Both classes "used AI for learning". Only one used it to make pupils do the learning. this seems so simplistic as I write it but we have to remember a tool is only as good as the person using it, we can train our teams to use the tool for best effect. That is the distinction this report helps us make, and it matters most with younger children, who are still developing the self-regulation to recognise when a shortcut is costing them something (although I am not yet convinced that many humans will be able to resist the shortcut in any event).

Access is not ability

I find it hard not to draw a parallel with how AI models themselves are built. A model becomes capable through an enormous volume of practice: predicting, being corrected and adjusting, over and over. The mechanisms are very different from a child's learning, but the principle is instructive. Capability comes from practice, and it does not transfer to whoever is holding the device. A pupil with access to a capable system has not acquired its capabilities.

Writing is where this matters most to me. A meta-analysis of 56 experiments found that writing about subject content improved learning in science, social studies and mathematics, particularly when pupils had to select, organise and explain information in their own words. That supports the value of writing as a learning activity.

If AI produces the explanation, the pupil has a useful text to read but has missed the chance to construct it. Writing skill itself develops through years of composing, revising and persevering. Each time we replace that practice, we reduce the experience through which the skill grows, and we take away opportunities for pupils to discover what they think and where their understanding is incomplete.

You could argue that these skills will matter less because AI can do so much of the writing for us. Even as a technology evangelist, I would disagree. Writing contributes to learning and memory, and it is how we learn to organise an argument, communicate precisely and form an independent view.

That does not mean preserving every difficulty. Appropriate support can make a task accessible and help a child persist. The judgement is whether we are removing a barrier to learning or removing the thinking and practice that the learning requires.

Three questions to ask before any AI trial

For school and trust leaders, this has direct implications for investment and implementation. "Increasing AI use" is too broad an objective to guide a useful evaluation. Before approving any trial, I would ask:

1. Which learning process is this meant to improve? Retrieval, feedback, explanation, reasoning? Name it.
2. What will pupils actually do? A retrieval tool should show how pupils retrieve information themselves. A reasoning tool should show how pupils develop and explain their own reasoning. Generating questions or explanations is a capability of the software; whether pupils learn more from using it is what we need to establish.
3. What evidence will show it worked, with the support removed? For a practice tool, look at what pupils can remember and explain later without it. For a tool claiming to develop critical thinking, look for evidence that pupils can evaluate unfamiliar material and justify their decisions independently. Those checks should reflect what was taught and include appropriate access arrangements.

I would also want to know who benefits and who does not, including pupils with SEND, disadvantaged pupils and those learning English as an additional language. Alongside that data, I would want qualitative feedback from colleagues: where the support helped, where it supplied too much of the answer, and how much checking or correction it needed.

Where I would start

My starting position is to explore AI for purposeful practice and consolidation, protect the effort pupils need to acquire knowledge, and require convincing evidence before accepting claims about higher-order thinking. Above all, we need to distinguish what technology can produce for a child from what that child needs to practise becoming capable of doing.

---

Sources: Kendeou and colleagues (2026), Ways Generative AI Can Support and Threaten Learning in K-20 US Education: Expert Consensus Report; Graham, Kiuhara and MacKay (2020), The Effects of Writing on Learning in Science, Social Studies, and Mathematics: A Meta-Analysis; Department for Education (updated January 2026), Generative AI: product safety standards. The Year 4 scenario is hypothetical. The AI training analogy and the proposed evaluation approach are my own reflections.