Who is Doing the Thinking?

1. Get pupils thinking before you tell them

In physics, I often ask pupils to predict what will happen before a demonstration.

That small act of committing to an answer changes the activity. If pupils simply watch, they can remain passive. If they predict first, they have something to compare with what actually happens. If they are wrong, there is now something to explain.

Research on the generation effect shows that information people generate themselves can be remembered better than information they simply receive. A meta-analysis of 86 studies found a clear overall advantage for generation over reading (Bertsch et al., 2007).

Even an unsuccessful first attempt can sometimes help. Research on pretesting has shown that attempting questions before studying the material can improve subsequent learning, despite many initial answers being wrong (Richland, Kornell & Kao, 2009).

Productive failure research points in a similar direction. Sinha and Kapur's meta-analysis of 53 studies found an overall advantage when problem solving preceded instruction, although the effect depended on factors including age, task and implementation (Sinha & Kapur, 2021).

This is not an argument for leaving pupils stuck.

Give pupils something to think about before giving them the thinking.

Crouch and colleagues demonstrated this in physics. Students who simply observed classroom demonstrations showed little improvement in conceptual understanding, while those who first predicted what would happen learnt more from what followed (Crouch et al., 2004).

The same principle applies to questioning. If we ask a difficult question and then answer it ourselves after two seconds, we have technically asked pupils to think without really giving them time to do so. Sometimes keeping the learner at the centre means being comfortable with a few seconds of silence.

Jamie Thom makes a similar point in Slow Teaching: important questions need thinking time. The aim is not to slow everything down, but to be more deliberate about the moments where thought matters.

AI makes it incredibly easy to skip the first attempt. That is precisely why we need to protect it.

2. Don't remove the useful struggle

Educational technology is often sold on efficiency. Usually, that is a good thing.

We do not want pupils wasting mental effort because the Wi-Fi is unreliable, instructions are confusing or a resource is difficult to access.

But learning presents an awkward problem:

Some inefficiency is useful.

Trying to recall something is slower than looking it up. Selecting the correct method is slower than asking AI. Constructing an explanation is slower than copying one. Finding your own mistake is slower than being shown the correction.

Yet these may be exactly the processes we want pupils to undertake.

This is where Thom's broader argument in Slow Teaching resonates with me: slow down the things that require thought. Slow does not mean making every lesson longer or less efficient. It means deliberately investing time where thinking matters.

I find it useful to distinguish between productive friction and unnecessary friction.

The goal is not maximum difficulty. It is maximum useful thinking.

This is one reason I have always liked the motto used by Isaac Science:

"Isaac. You work it out."

It is deliberately simple, but it captures something important about learning. Getting stuck, trying an approach, discovering that it does not work and trying something else can all be part of becoming a better problem solver. The educational value is not simply in eventually having the correct answer. It is in doing the thinking required to reach it.

That gives educators a useful question whenever technology makes something easier:

What exactly is being made easier?

If it removes administration or improves accessibility, excellent. If it removes the need to retrieve, decide, reason or explain, we should be more cautious.

3. Scaffold thinking, don't replace it

Good teaching is full of support. We model, explain, prompt, provide worked examples and break complex tasks into manageable steps.

The issue is not whether support exists. It is what that support gets the learner to do.

That also matters when we model. A worked example can be extremely helpful, especially for novices, but the value lies in more than displaying a polished final answer. Good modelling makes the reasoning visible: why this method was chosen, what alternatives were rejected and what an expert notices that a novice may miss.

In other words, we should model the thinking as well as the product. The support can then be gradually withdrawn so that pupils increasingly carry out that thinking themselves.

Recent AI research makes this distinction unusually visible

Bastani and colleagues studied nearly 1,000 secondary-school mathematics pupils using different forms of GPT-4 support. Pupils with access to a relatively unrestricted AI system performed much better while they could use it. Once the AI was removed, however, they performed 17 per cent worse than the control group (Bastani et al., 2025).

A second system was deliberately designed to behave more like a tutor and make it harder for pupils simply to obtain solutions. The negative effect on subsequent independent performance was largely mitigated.

There is promising evidence in the other direction too. In a 2025 randomised controlled trial involving 194 university physics students, Kestin and colleagues found that students using a carefully designed AI tutor achieved greater learning gains than those in the comparison active-learning class while spending less time on the task (Kestin et al., 2025).

The question, then, is not "Is AI good or bad for learning?" It is:

What is the AI getting the pupil to do?

An AI that simply provides the answer may remove thinking. One that asks questions, gives graduated hints, diagnoses errors or generates new practice may increase it.

And effective AI use still depends on knowledge. Pupils need enough subject understanding to recognise when an explanation is implausible, identify where they are stuck and judge whether the help they receive is useful. AI literacy is not an alternative to subject knowledge. In important ways, it depends upon it.

4. Make pupils do something with their mistakes

Retrieval practice rightly receives considerable attention in schools. Trying to bring information back to mind can strengthen later retrieval and slow forgetting (McDermott, 2021).

But retrieval can easily be reduced to:

Question → answer → mark → next question

The more interesting part is what happens when the pupil is wrong. Butler and Roediger found that corrective feedback after testing increased later correct responses while reducing retention of incorrect alternatives (Butler & Roediger, 2008).

For me, this suggests a richer cycle:

Diagnosis matters. A pupil who gets a physics calculation wrong might have forgotten the equation, selected the wrong equation, misunderstood the concept or simply made an algebraic error. Those are different problems and require different responses.

This diagnosis also develops an important element of metacognition. Pupils become better at recognising what they know, where they are struggling and what they need to do next, rather than simply receiving a mark and moving on.

The pupil then needs to reattempt. Reading a correction is not the same as performing the reasoning again.

Finally, they need to revisit it later. Learning takes time. Research on memory consolidation, including the role of sleep, reinforces the point that learning continues beyond the initial encounter (Rasch & Born, 2013).

We do not need complicated neuroscience-based teaching strategies from this. The classroom implication is simpler:

Feedback should trigger another attempt, and successful performance once should not be mistaken for secure learning.

5. Practise the weakness, not just the topic

When a pupil struggles, we often say they need "more practice".

But more practice at what?

Consider three pupils struggling with Newton's second law.

All three appear to need more work on "forces", but they do not have the same problem.

Research on deliberate practice emphasises focused activity aimed at improving specific aspects of performance, accompanied by feedback and repeated opportunities to improve (Ericsson, 2008).

So perhaps the more useful question is: what specifically can this pupil not yet do? Then practise that.

This is one area where AI could become genuinely useful. Not because it can generate endless worksheets, but because it may help identify patterns in errors and make targeted practice easier to provide.

The teacher remains responsible for the judgement, but technology may help make responsive practice more practical.

6. Find out what remains when the support disappears

AI creates a serious risk of false mastery.

A pupil produces a sophisticated essay and we infer sophisticated understanding. They solve difficult mathematics with AI and we infer mathematical capability. They follow a fluent explanation and feel that they understand it.

But what remains when the support disappears?

This is why the Bastani study matters. Pupils looked more successful while AI was available even though subsequent independent performance deteriorated in the less constrained condition (Bastani et al., 2025).

Teachers therefore need regular opportunities to remove the scaffold.

This does not mean everything should become a closed-book test.

Daniel Susskind offers a useful parallel with the arrival of calculators: "teach both, test both." Pupils should learn to think, reason and solve important problems independently, but they should also learn to use AI effectively when it is appropriate. An AI-assisted piece of work and an independent piece of work can both be valuable, but they provide evidence of different capabilities (Susskind, 2026).

Thom's emphasis on planning for learning over time is useful here too. Good teaching cannot be judged only by how successful pupils appear during a single lesson. What matters is what has become durable enough to be retrieved, applied and built upon later.

For homework in particular, the question is increasingly not simply "Has this been completed well?" but:

What evidence do I have that the pupil can do the underlying thinking?

7. Choose the tool to fit the learning

I lead on digital strategy in school, but I do not think the aim of digital education should be to maximise technology use. That would confuse the tool with the goal.

Sometimes the best tool will be a laptop. Sometimes an AI tutor. Sometimes a mini-whiteboard. Sometimes a discussion. Sometimes a blank sheet of paper.

Handwriting illustrates the point. Claims that handwriting is simply "better for learning" than typing are too strong. A 2024 meta-analysis found a small achievement advantage for students who took and reviewed handwritten notes, although students who typed recorded substantially more information (Flanigan et al., 2024). Another systematic review and meta-analysis found no reliable overall advantage for either method when digital distraction was controlled (Voyer, Ronis & Byers, 2022).

So the lesson is not paper good, screens bad. It is:

Choose the medium according to the thinking you want pupils to do.

The same applies to debates about screen time. An hour passively scrolling, an hour constructing a simulation and an hour receiving carefully scaffolded tuition may all count as an hour of "screen time", but educationally they are very different activities.

Typing is excellent for editing, collaboration and reorganising information. Handwriting can be useful when pupils need to select information, annotate, manipulate equations, sketch relationships or retrieve without the distractions of a connected device.

AI can be valuable when we want hints, targeted practice, alternative explanations or critique. It may be much less valuable when the point of the task is to practise the thinking that the AI would perform. And where AI is deliberately allowed, perhaps the opportunity is not simply to make the existing task easier, but to raise the ambition of what pupils are asked to do.

A 1:1 device is not a pedagogy. Neither is AI.

Their educational value depends on what pupils are asked to do with them.

Who is doing the thinking?

None of these principles began with generative AI.

We already knew retrieval mattered. We already knew generation and prediction could support learning. We already knew feedback needed to lead to action. We already knew scaffolds should eventually be removed and that practice should target what pupils cannot yet do. We also knew that learning often benefits from slowing down at the moments where thought matters most.

AI has not overturned learning science. It has made understanding it more urgent.

For the first time, pupils have easy access to technology capable of performing a remarkable amount of cognitive work on their behalf. That creates risks. It also creates genuine opportunities for tutoring, feedback, diagnosis and targeted practice.

The challenge for educators is therefore not to maximise or minimise AI use. It is to decide which thinking should remain with the learner.

If we keep asking those questions, we have a much better chance of using technology to improve learning rather than simply improve the appearance of learning.

That, for me, is what it means to keep the learner at the centre.