AI in Education #3: Five things I learned from Adam Boxer about why AI will not revolutionise your classroom

This is the third in a series of AI in Education specials, where I speak to the world’s leading experts about how AI is changing teaching and learning.

Adam Boxer is one of those annoying people who makes you think harder than you wanted to. He is a science teacher at The Totteridge Academy (TTA) in London, the co-creator of Carousel Learning, a best-selling author… and I believe he also has a podcast?

It was an absolute pleasure to have Adam back on the show. He is, as I suspected before we sat down to talk, more sceptical about AI in education than most people I speak to. So, let me share five practical takeaways from our conversation that I think every classroom teacher and school leader will find interesting.

1. Student-facing AI tools are a terrible idea right now

I expected Adam to be cautious. I did not expect him to be this direct. His position is clear: he is completely against any student-facing AI tool at this point in time, and probably will be for quite a while.

Why? Because Adam watches students doing homework. And what he sees is what any teacher who has spent five minutes in a real school would predict: students take the path of least resistance. Every single time.

I pushed back on this. I used Adam’s own argument against him – the one he has been making for years about integrating homework with classwork. My pitch was simple: let students use AI for homework, then test them in class without technology. If they ace the test, they used the AI well. If they bomb it, they did not, and you can support them to use it more effectively. Problem solved, right?

Adam was not buying it. He made a point that really stuck with me: even if you tell students to use AI responsibly, you have broken the seal. They no longer see using AI as a problem. And while they might use your carefully constrained AI tutor in maths, they will just open ChatGPT in another tab for history and English. As Adam put it: “You need to communicate to students that using AI is dumb and it is a waste of their time and will rot their brain.”

Strong words. But then he went further. He referenced something Michael Pershan has written: any tutor that the tutee can control is not a tutor. The moment you open a chat interface, the student becomes the agent. They say “give me a hint.” Then another hint. Then another. And before you know it, the AI is doing all the thinking.

I am still not entirely sure Adam is right on this one. I think there is a version of AI-supported homework that could work for some students with the right incentives. But his core point is hard to argue with: we have no evidence that putting students in front of AI tools leads to learning, and we have plenty of reasons to think it will not.

2. AI can help with marking – but the sweet spot is not where you think

Here is where Adam surprised me. Given how sceptical he is about everything else, I expected him to dismiss AI-assisted marking out of hand. He did not.

Adam recently marked his Year 11 mocks. He described what he learned from the process: students were not structuring their six-mark answers properly, they were not referring to graph data when the question asked them to, and there was a topic on nitrogen in the atmosphere that he realised he had not taught well enough. That knowledge came from sitting with the papers.

But he is also honest about the reality. Not every teacher gets those insights from marking. Some teachers just plough through the papers without really thinking about it. And as I pointed out, by the time I finished a marathon Sunday marking session, I was so exhausted I could not spot the trends anyway.

So where is the sweet spot? Adam thinks it is this: AI handles the bulk of the tick-cross marking, then produces a report that the teacher is obliged to review and interact with. The report flags trends, highlights specific student work, and identifies misconceptions. The teacher still engages with the substance – but without spending hours on the mechanical stuff first.

He gave a shout-out to Josh Perry and Smart Grade, whose science marking tool he described as “genuinely amazing.” And he mentioned that the project he is working on at Carousel takes a similar approach: AI gives feedback on students’ retrieval practice work, but only in situations where teachers were not already marking it, and is built in a way that forces the teacher to actually do something with the information.

I think Adam is right that there is a sweet spot here. And I think his framing is helpful: AI should step in where teachers were not engaging with the work anyway, not replace the marking that was already building their professional knowledge.

3. AI-generated resources might make the average teacher better and brilliant teachers extinct

This was the part of our conversation that made me most uncomfortable, because I think Adam might be right about something I do not want to be true.

Here is his argument. Imagine a normal distribution of teacher quality. AI-generated resources will eliminate the bad tail. A teacher who would have produced a terrible worksheet can now type a prompt into ChatGPT and get something passable in five seconds. That is a genuine improvement for those students.

But what about the other end? Adam has written over 100,000 questions throughout his career. He built booklets for every chemistry module. That intellectual labour is what made him the teacher he is today. If the next generation of teachers never writes a single question because ChatGPT does it for them, where do the brilliant resource creators come from?

He put it to me bluntly: “You would not exist.” And he is right. I have written God knows how many multiple-choice diagnostic questions. That process – the thinking, the refinement, the understanding of assessment theory that came through doing it thousands of times – is not something you get by typing a prompt.

I pushed back. I told Adam about my recent experience using Claude as a thought partner for question writing aligned with psychometric standards. I write the quiz, then ask the AI to evaluate it as a psychometrician would. It asks me why I put distractor B there, why question two comes before question four. I answer, it gives feedback, and I learn. I think I am getting better at question writing because of that process.

Adam conceded this might work for experienced teachers at the top end. But he made a point I could not argue with: “You have been in schools long enough to know that is not how your bottom 50% of teachers are going to be using it. They are going to go, can you give me 10 multiple-choice questions about Pythagoras, and that is what they are going to use.”

He even tried it himself. He asked ChatGPT to make a worksheet on thin-layer chromatography for A-level chemistry. The output was, in his words, trash. His students spotted the errors. He has not used it since.

So what is the solution? Adam’s approach at TTA is pragmatic: use AI if you want, but if I come into your lesson and the worksheet is rubbish, I do not care where it came from. And I will ask you pointed questions. Can you generate five more questions on the spot if students are stuck? Do you know why the example sequence is ordered that way? If the answer is no, we have a problem.

4. AI coaching for teachers is promising but terrifyingly easy to get wrong

Adam recently interviewed researcher Raj Chande on his podcast. Raj’s team is exploring whether AI can analyse lesson transcripts and give teachers feedback. Adam’s eyebrow was arched throughout.

The basic idea: record a lesson, transcribe it, and feed the transcript through an AI trained by expert mentors to recognise effective teaching moves. The AI spots things like whether the teacher put the student’s name at the beginning or end of a question, how long the wait time was, and how many boys versus girls were asked questions.

Adam sees genuine potential in some of this. Wait time placement, name-at-the-end questioning, gender distribution – these are things a human observer struggles to track in real time, and an AI working from a transcript could measure them well. He would use such a tool to run 15 lessons through the system and look for department-wide trends. That is strategic. That is useful.

But here is where he gets worried. A transcript is a tiny fraction of what happens in a lesson. It misses physicality, tone, body positioning, what students are doing. And the AI’s training data is – as Adam put it – “worse than trash” when it comes to education. How do you stop the AI from telling teachers they should have used thinking hats or learning styles?

He gave an example that I thought was brilliant. Imagine the AI says: “You spoke for four minutes without asking any questions.” Sounds like useful feedback, right? But Adam has no problem with a teacher talking for ten minutes without a question, provided students are still paying attention. And when the teacher does pause to ask a question, Adam wants it to be the shortest, sharpest, easiest question imaginable – just to check students are still with them. The AI, trained on Rosenshine’s principle that effective teachers “ask lots of questions,” might flag that easy question as too simple. And the teacher, trusting the feedback, starts asking harder questions mid-explanation, which is exactly the wrong thing to do.

Adam’s fear is not the technology itself. It is what happens when the technology escapes the hands of the people who understand teaching well enough to interpret it. And he is already seeing it. People have contacted him about receiving AI-generated feedback on their teaching that they did not ask for and do not agree with.

His summary: “I trust Raj and his team. I do not trust everyone else.”

5. AI will not revolutionise education – because nothing ever does

I saved the best for last. When I asked Adam to make a prediction about the future of schools, he did not hesitate.

“AI is not going to revolutionise education. It is not going to happen. And what is the proof of this? That nothing ever does.”

He rattled off the list: Encarta, Google, Wikipedia, interactive whiteboards. Every one of them was going to transform the classroom. None of them did. Why? Because schools are not nimble, agile organisations. They are tankers. They change slowly, incrementally, over years.

Adam’s view is that England’s education system is getting better – but slowly. Better behaviour standards. Workloads trending down. Growing awareness of cognitive science and retrieval practice. Student outcomes ticking up one percent, one percent, one percent. AI might feed into that gradual improvement. But it is not going to cause a spike.

He invoked something called the Lindy Effect – an idea from Nassim Nicholas Taleb that the longer something has been around, the more confident you can be that it will endure. Schools have been around for over a hundred years. The model of one teacher and thirty students works. It serves important societal functions. And as Adam pointed out, ed-tech futurists who dream of students learning on screens all day have clearly never worked in a tough school. “Have they never seen an IT lesson? It is carnage.”

He also made a point about the gap between aspiration and reality that I think deserves to be quoted in full: “There is such a big chasm between what people on LinkedIn think will happen in schools and what schools are actually really like. You go to a typical school, and there will still be posters on the wall from ten years ago.”

I found Adam’s prediction oddly reassuring. Not because I think AI is useless – I use it every day, and I think it has genuine potential in the areas we discussed. But because the hype can be overwhelming, and it is easy to feel like you are falling behind if your school has not adopted the latest AI tool. Adam’s message is simple: you are probably fine. Focus on the things that matter – clear explanations, checking for understanding, retrieval practice, strong behaviour – and let the AI stuff settle before you rush to adopt it.

Over to you

This was a brilliant conversation, and I have only scratched the surface here. You can listen to or watch the full episode on the Mr Barton Maths Podcast. And if you have not already, check out Adam’s podcast, They Behave For Me (the UK’s second-best education podcast) and the work his team is doing at Carousel Learning.

This is the third in a series of conversations I am having with leading thinkers about AI in education. The first, with Daisy Christodoulou, is here. The second, with Barbara Oakley, is here. Next up: Carl Hendrick, who has been spending time at Alpha Schools and might just disagree with everything Adam has said.

I would love to know: what is your school doing about AI? Are you letting students use it? Banning it? Somewhere in between? And do you think Adam is being too sceptical, or not sceptical enough?

Let me know. And as always, I could be wrong about all of it.

Craig

🏃🏻‍♂️Before you go, have you… 🏃🏻‍♂️

… checked out my brand-new book series: The Tips for Teachers guides to…

And checked out my all-new, ad-free, Mr Barton Maths website