AI in Education #6: Five Things I Learned from Dan Meyer about Why AI Should Serve Teachers, Not Replace Them

This is the sixth 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.

Dan Meyer is one of those rare people in education who builds the technology and criticises it. He is a former secondary maths teacher from the US, now at Amplify (formerly Desmos), where he led the development of the company’s first classroom AI feature. He is also one of the most prominent sceptics of AI tutoring hype. Alongside his day job, he spends his weeks visiting classrooms and teaching lessons himself, which he describes as “a fantastic inoculation from some of the more prominent AI psychosis.”

I last spoke to Dan on this podcast a decade ago. A lot has changed. But his instinct – that the sophistication of teaching is routinely underestimated by people building technology – has not.

It was an absolute pleasure to have Dan back on the show, and I wanted to share five takeaways from our conversation that I think every classroom teacher and school leader will find interesting.

1. AI tutors have failed to deliver – and anyone who has actually tutored a student could have predicted that

Two weeks before we recorded this episode, Sal Khan of Khan Academy admitted that Khanmigo, their personal AI tutor, is not doing what they thought it would do. Dan was not surprised. As he put it: “I was not alone, but I was naming this early.”

His argument is simple. If you have ever sat beside a student and helped them through a problem, you know the decisions involved are constant and subtle. When do you intervene? When do you hold back? When does a student need encouragement, and when do they need to be told: “No, keep going, you are closer than you think”? These are not just cognitive decisions. They are social ones. They draw on everything you know about the student – what happened in yesterday’s lesson, what is going on at home, how they respond to different kinds of pressure.

Dan pointed to Khanmigo’s struggle with something as basic as when to offer help. First, they had it off until the student clicked it. Nobody clicked. Then they triggered it when a student clicked into the answer box and clicked out. Now it is just always on. These are the largest companies in ed-tech, and they cannot solve the timing problem.

I pushed back. I pointed out that in the UK, we have a retention and recruitment problem, resulting in non-specialist teachers covering Key Stage 3 classes, kids having two or three different maths teachers in a week, and that we have to compare AI tutors against that reality, not against some utopian vision of a super-skilled, super-experienced teacher who intimately knows every child.

Dan agreed that the problems are real. But then he said something that gave me pause. The problems I was naming – behaviour, burnout, teacher shortages – are problems that demand solutions. Some of those solutions could be technological. But they are also political: paying teachers more, reducing class sizes, supporting families so their kids come to school better nourished and better prepared. Those solutions have historical precedent and good odds of success. They just cost money. And for some reason, they are deemed out of bounds before the conversation begins.

2. The AI features that actually work make the teacher-student relationship stronger, not weaker

So if Dan is sceptical about AI tutors, what does he think AI should be doing in classrooms?

He showed me. Amplify has built a feature called Discussion Moments. Here is how it works. In a lesson on integers, students manipulate a submarine up and down by adding floats or subtracting anchors. Towards the end, they are asked: here is a star at five – what combination of floats and anchors would reach it? Try to come up with something no one else would think of.

Students type their responses. The teacher sees all 35 on screen. And here is where the AI comes in: it processes those 35 free-text answers, selects three that successfully reached the star and one that miscalculated, and presents all four on a discussion slide with a narration for the teacher. The teacher reads: “Which one of these is not like the others? Are they all correct? Talk amongst yourselves.”

The AI does the language processing. The teacher does the work of building relationships around mathematical content.

This connects to something that has come up again and again in this series. Daisy Christodoulou argued for 10% human judgements and 90% AI in assessment. Simon Woodhead described a human-in-the-loop model for question writing where the teacher’s corrections feed back into the model. The pattern is the same: AI works best when it mediates human relationships, not when it replaces them.

I asked Dan the obvious question: Why not cut the teacher out? Why not have the AI select those four responses and show them directly to the student, with personalised follow-up questions? His answer was the line from our conversation that I think will stick with me longest…

(cliff-hanger!)

3. Students do not care about showing off for an LLM

… Dan put it bluntly: Students do not care about bringing their best work to an AI chatbot. They do care about their classmates seeing them. They care when the teacher says, “Ooh, this is wrong, but very interesting.” That has an effect – socially, on every single person in the classroom – that, day in, day out, contributes to a class where students think: I need to bring my best here.

And you can only teach a student to the degree that they are willing to offer you their thinking. If they are not expressing their ideas – sincerely, not just gaming the system for a hint – then it does not matter how good the AI is. It is game over.

I pushed back with evidence from our Eedi study: some students told us they were more willing to ask questions of an AI than of their teacher, because they were less embarrassed. Dan conceded that there are socially toxic classrooms where that might matter. But his broader point is about the vast majority of students and the vast performance of AI in 2026: the social incentive structure of a real classroom is doing work that no chatbot can replicate.

This is where Dan parts company with Carl Hendrick. When I pushed Carl on the Alpha School model – asking whether removing the social element of 30 kids learning together would kill motivation – Carl argued it was still social, just a different kind. The guides have one-to-one conversations. The students are held accountable through their scores. But that feels like a very different game from the one most schools are playing, and a much easier one when you have pre-selected for motivated parents who are paying $40,000 a year.

Dan had a good line on this: “This is someone who plays cricket saying about football, ‘you are not playing cricket right.’ These are different games. And I think yours is a lot easier.”

4. Teachers are in a much better position than students to learn from AI

Here is where Dan’s scepticism starts to lift. When it comes to teachers using AI for their own professional development, he is genuinely enthusiastic.

Why the difference? Teachers have a different level of literacy. They have a stronger self-conception as learners. They have different expectations for how they interact with software. And crucially, they can evaluate the output in ways most students cannot.

Dan is interested in AI generating practice problems aligned to a curriculum, creating isomorphic test questions so students can retake assessments, and helping teachers adapt existing resources for their specific classes. He also sees real potential in AI analysing lesson transcription data to offer teachers feedback on their facilitation – an area he is actively exploring at Amplify.

But he made a crucial point about teacher variability. Someone said that, in software engineering, the best engineers will be multiplied by 10x by AI, while the least experienced will be multiplied by 0.9. Dan thinks the same may be true for teaching. An experienced teacher can catch when the AI produces a word problem where the units do not match, or the context is nonsensical. A novice teacher might not.

This connects to a thread running through the whole series. Adam Boxer worries that if teachers stop writing their own questions, the next generation of brilliant resource creators will never develop. Barbara Oakley argued that AI is a brilliant thought partner for experienced teachers, but it is not a curriculum. Dan agrees with both: AI should be a multiplier of teacher expertise, not a substitute for it. And the products that try to produce identical outcomes regardless of teacher experience are, in his view, sidelining one of the most effective resources in a student’s development.

5. 95% of students are excluded from the success stories – and that should bother us more than it does

Dan reminded me of Laurence Holt’s 5 Percent Problem. If you analyse the efficacy reports from personalised learning companies – the ones that claim their AI tutoring platform produces measurable gains - you typically find that the positive results only emerge after excluding what appears to be 90–95% of the study population. The students who did not engage enough, who did not complete enough sessions, who dropped out.

If the technology only works for 5% of students, that still leaves 95% unserved. And the 5% it works for are likely the students who were already most motivated and most capable of self-directed learning.

Our CEO, Ben Caulfield, is a massive fan of Dan’s work, and when Dan wrote the post highlighting this 95/5 split, it became a mantra at Eedi: If we are not serving the 95%, what is the point?

I think this is the most important question anyone in education technology should be asking. Not “Does this work in a controlled study with motivated participants?” but “Does this work at scale, with real kids, in real classrooms, including the ones who are on the floor with undiagnosed ADHD?” Because that is the population we are trying to serve.

Bonus: AI is all potential energy and no kinetic – so why do we keep talking about it?

I asked Dan the question I ask all my guests: What question about AI have I not asked you that I should have? His answer was different from anyone else’s in the series.

He asked: Why do we all talk about AI so much?

Relative to other technological innovations – the steam engine, the combustion engine, vaccinations, the telegraph – AI has not yet transformed the day-to-day experience of most people. It is, as Dan put it, all potential energy and no kinetic. That is not to say students are not using it to cheat, or that there is no job displacement. But the gap between the hype and the reality is vast.

So why does AI keep getting pushed forward as the solution to problems in education – novice teachers, poor behaviour, burnout – when there are proven solutions with historical precedent, like increasing teacher pay or decreasing class sizes?

Dan’s hypothesis: the success of AI would generate value for a small number of people who have a lot of interest in both media properties and AI companies. And so it does not surprise him that the media talks about AI quite a lot, and that people like you and me, further downstream, find ourselves talking about it too – instead of talking about the social and political solutions that are harder but more likely to work.

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 do check out Dan’s Substack, Math Worlds, which is one of the most thoughtful places to read about AI and education right now.

This is the sixth in a series of conversations I am having with leading thinkers about AI in education. Here are the others in the series:

1. Daisy Christodoulou: podcast and write-up

2. Barbara Oakley: podcast and write-up

3. Adam Boxer: podcast and write-up

4. Carl Hendrick: podcast and write-up

5. Simon Woodhead: podcast and write-up

I would love to know: where do you land after six episodes? Are you more optimistic about AI in education, more sceptical, or just more confused? And is Dan right that we should be spending more time talking about teacher pay and class sizes than about chatbots?

Thanks so much for reading.

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, with my new Topics page?