AI in Education #4: Five things I learned from Carl Hendrick about Why This Time It Really Is Different

This is the fourth 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.

In a rare moment of forward-planning, I scheduled Carl Hendrick right after Adam Boxer in my AI in Education podcast series, as I had a feeling they would offer differing views.

And they certainly did.

Adam came on the podcast and told me, in no uncertain terms, that AI will not revolutionise education, that nothing ever does, and that anyone who thinks otherwise has never set foot in a real school.

Carl came on one week later and said: This time it is different.

Same science of learning. Same respect for evidence. Same scepticism of ed-tech hype. Completely opposite conclusion.

Carl started teaching 20 years ago in an inner-city London state school. He did his PhD at King’s College, co-authored three books on evidence-informed teaching, and has spent the last decade trying to bridge the gap between cognitive science and classroom practice. He is now working with Alpha School in the US, building AI-powered learning apps grounded in what we actually know about how students learn.

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

1. Ed-tech has always been a failure – but this time, the science of learning is driving the design

Carl is not some wide-eyed ed-tech optimist. He has spent years arguing that ed-tech is a history of expensive failure. He said those exact words in our conversation.

So why does he think AI is different?

Because previous ed-tech – Encarta, interactive whiteboards, most so-called “adaptive” platforms – was built by people who did not understand the science of learning. Most of it was just digitised worksheets. Some platforms claimed to be adaptive, but they were not. Not really.

What Carl sees now is something new: AI systems that can gather high-resolution data on individual students – not just right or wrong, but latency, hesitancy, patterns of error across weeks – and then use that data to predict, adapt, and iterate in real time. And crucially, the people building these systems (or at least the ones Carl is working with) are grounding the design in what we know about encoding, retrieval, spacing, and the limits of working memory.

Carl put it bluntly: if learning is a biological process governed by the laws of biology and physics, then it is measurable. And if it is measurable, it is designable. Either you accept that, or you believe there is something supernatural about learning that puts it beyond the reach of science.

If we accept that cognitive science tells us real things about how students learn, then a system that can act on those principles at the individual level – adapting the spacing schedule, adjusting the sequencing, flagging misconceptions in real time – is qualitatively different from an interactive whiteboard.

2. Retrieval practice in schools is “the most lethally mutated thing in education”

I have spent a good chunk of my career banging on about retrieval practice. So when Carl described it as the most lethally mutated thing in education, I took note.

His argument is this: the principle of retrieval practice is sound. We know it works. But what he sees in schools is a version so watered down that it barely resembles the real thing.

Take vocabulary learning. Carl described a common pattern: students learn six words on a Monday, cram on Sunday night, score five or six out of six on Monday’s quiz, and then never encounter those words again in any structured way. The school ticks the retrieval practice box. But the words vanish within weeks because no one has thought about the spacing schedule, the optimal amount of forgetting before the first retrieval attempt, or how those words get encountered in rich context over time.

I saw versions of this in maths lessons every week. Do Nows that tested the same topic all week (as I wrote about in my Correcting green pen doesn’t work piece), exit tickets that were really just short-term memory tests, and retrieval schedules that looked systematic on paper but fall apart the moment a teacher is absent or a lesson runs over.

Carl’s view is that AI can solve this – not by replacing the teacher, but by doing the thing that no individual teacher can realistically do: track the optimal spacing interval for every student on every concept, predict when forgetting is about to tip from productive to destructive, and schedule the next retrieval opportunity accordingly.

Will this be significantly better than a teacher with a well-designed retrieval schedule or a good system, such as Adam Boxer’s Carousel Learning? I don’t know. But Carl’s point about the gap between the principle and the practice is one I cannot argue with.

3. Alpha School is not what you think it is

Before talking to Carl, my mental image of Alpha School was kids on laptops for several hours a day with no teachers in sight. That is not what it is… or at least, not exactly what it is.

Here is the model. Students spend about two hours in the morning working on AI-powered learning apps. If they hit their targets, they get the rest of the day back – sport, projects, being kids. And the results, Carl says, are strong.

Two things changed my mind about dismissing this out of hand.

First, chatbots are banned. Mobile phones are banned. This is not kids messing about on ChatGPT. The AI is doing the heavy lifting behind the scenes – sequencing knowledge, monitoring attention, gathering data – while the students are thinking hard about carefully designed questions and tasks. Carl described a “waste meter” that tracks time on task, latency, and whether students are actually engaged or just going through the motions.

Second, the human element remains. Alpha calls their staff “guides” rather than teachers, which will make some people wince (it made me wince). They are not necessarily subject experts - the AI is the subject expert. But what these guides actually do is have one-to-one conversations with students based on the data the system generates. They look at where a student is struggling and ask: is it a knowledge gap, a motivation issue, a confidence problem? As Carl pointed out, the paradox is that students at Alpha may get more individual tutor time than students in a traditional classroom of 30.

Now, I need to be careful here. Alpha charges upwards of $40,000 a year. The data on outcomes is not public. And as Adam pointed out in the previous episode, there have been serious allegations about content scraping. I am not saying this is the future of education. But I am saying it is more interesting than I expected, and the model deserves scrutiny rather than dismissal.

4. Most teachers are terrible at curriculum design – and that is not their fault

Carl argued that curriculum design is one of the hardest things in education. Decomposing a rich domain of knowledge into constituent parts, figuring out the right sequence, knowing what prior knowledge students need before they encounter a new idea – this is brutally difficult work. Most teachers arrive at decent sequencing through years of trial and error. Many never get there at all.

And here is where Carl thinks AI can help. Not by replacing the curriculum designer, but by capturing the knowledge of the best teachers – the ones who understand, as Christine Counsell puts it, the “semantic gravity” of their subject – and using it to train systems that can sequence knowledge in ways that most individual teachers simply do not have the time or expertise to do from scratch.

I have long wrestled with the idea of scripted lessons and centrally designed curricula. For the first 12 years of my career, I would have hated the very notion. But the older I get and the more lessons I watch, the more I think most students would be better off with a brilliantly designed curriculum delivered by an average teacher than with an average curriculum made up on the fly, even by a good one.

Carl’s argument is that AI can make that brilliant curriculum more available, more adaptive, and more responsive to individual students than a textbook or a scheme of work ever could. I think he might be right about that. But Adam’s worry from the previous episode still nags at me: if teachers stop doing the intellectual labour of curriculum design, where does the next generation of brilliant curriculum designers come from?

5. The science of learning is about to get a lot more interesting

This was the part of our conversation that got Carl most animated. And I think it is the most important takeaway.

Here is the problem with the science of learning as it stands: most of what we know comes from studies of 40 postgraduate students in a lab. The signal-to-noise ratio, as Carl put it, is low. Tom Perry’s review from five years ago made this point sharply – we still have almost nothing on retrieval practice in an actual classroom. We know some spacing is better than none. We do not know what the ideal spacing schedule looks like for a Year 9 student learning about ionic bonding on a Tuesday afternoon.

But what if millions of students were using AI-powered learning platforms, and every interaction generated data? What if we could track not just whether a student got a question right, but also how long they hesitated, what types of errors they made, and how their performance changed across days and weeks?

Carl’s prediction is bold: in the next two to five years, we are going to discover things about learning that we have been trying to figure out for decades. The platforms will generate the data. The AI will find the patterns. And the science of learning – which has been stuck on generalities for a long time – will finally get the resolution it needs to be useful at the level of individual students in real classrooms.

I find this exciting and terrifying in equal measure. Exciting because I have spent 20 years trying to apply cognitive science to maths teaching, and I know how frustratingly vague the guidance can be. Terrifying because, if Carl is right, the implications for the teaching profession are enormous. Not tomorrow. But sooner than most of us think.

Boxer vs Hendrick: where does that leave us?

I cannot write this piece without addressing the elephant in the room. Adam Boxer says AI will not revolutionise education. Carl Hendrick says it already is. They are both smart, both grounded in cognitive science, and both care deeply about what happens in classrooms. So who is right?

Adam is right that schools are tankers, not speedboats. He is right that every piece of ed-tech in history was supposed to change everything and changed almost nothing. He is right that the gap between what people on LinkedIn think will happen and what actually happens in schools is enormous.

But Carl is right that this technology is qualitatively different from an interactive whiteboard. He is right that the science of learning needs better data than we currently have. And he is right that pretending nothing has changed is its own kind of delusion.

My best guess? The truth is somewhere in the muddle between them. AI will not replace teachers. It will not turn every classroom into Alpha School. But it will – slowly, messily, unevenly – change what it means to teach and what it means to learn.

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 Carl’s Substack, where he writes about new papers and ideas every week.

This is the fourth 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. The third, with Adam Boxer, is here.

I would love to know: has this series changed how you think about AI? Are you more optimistic, more sceptical, or just more confused than when we started? And who should I speak to next?

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