Here are three blog posts I found interesting this week:

1. Let me convince you to try writing a script by Alex Blanksby

The idea of lesson scripts always splits the crowd. Alex argues that the case against scripts collapses once you stop picturing a 60-minute monologue and start picturing 30-second chunks slotted together, re-sequenced, thrown out and rebuilt as the lesson unfolds.

2. The Medium Is the Message by Dylan Kane

Dylan argues that the medium is never neutral, and offers “videogamification” as his counterexample. Students bring video-game habits to any screen, chasing completion rather than learning. On his own maths fact website, some worked out how to game it. How did Dylan address the problem?

3. Motivation — no, it’s not 90% of learning, but it does matter to an extent by Daniel Muijs

Daniel argues that motivation’s effect on attainment is real but modest, and that the research is muddied by the fact that “motivation” means half a dozen different things across studies. The directional finding is not the one most people assume.

Have a great weekend!
Craig

🏃🏻‍♂️Before you go, have you… 🏃🏻‍♂️

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

And checked out my Ultimate Retrieval Tool page on my Mr Barton Maths website

This interview was not supposed to happen.

I was booked to talk to Manu Kapur about productive struggle, but Manu did not show up (perhaps Manu was trying to give me a concrete experience of productive struggle by forcing me to think on my feet and find a guest at the last minute). So, I messaged Kris Boulton and asked if he fancied stepping in. He did, and within about three minutes, he had invented a format for us.

Kris runs Unstoppable Learning, where he works with schools and teachers to improve the quality of their instruction, particularly in how to sequence the examples they use to teach a concept. He had been sent a teaching and testing sequence on the volume of a prism, built by an early-career teacher he is working with. Here it is:

Take a few moments to look at it.

• What do you like?

• What would you change?

I think this sequence is good. The variation is controlled, with only one feature changing at a time. The teacher has generalised from the start, using “area times length” rather than a cuboid-only formula. Two teacher examples, then straight over to the students. I would have been delighted to produce that ten years into my career, and it was made by someone in their first couple of years.

Then Kris rebuilt it:

Again, before we get into the nuts and bolts, two further questions:

• What differences can you see?

• Why do you think Kris made these changes?

In our conversation, I focused on the five differences I think travel furthest - the ones you could apply to almost any sequence you plan, whatever the topic - and quizzed Kris all about them.

Enjoy!

1. Start with the general case, not the cuboid

In virtually every single resource on the volume of a prism, the first prism chosen is the cuboid:

This makes perfect sense as the cuboid is the simplest prism, so a natural starting point.

The trouble is that a cuboid is the one prism where the rule behaves oddly.

In a cuboid, every face is a rectangle. So it is not obvious which face is “the cross-section”. In fact, you can find its volume in three different ways, by selecting any of the three faces and multiplying by the length perpendicular to it. Every other prism has one face that visibly stands out as different from the rest. In a triangular prism, the triangle is the only non-rectangular face, so the question “which area do I use?” answers itself straight from the diagram.

That is why Kris starts with a triangular prism, and saves the cuboid until the very end.

And when the cuboid finally arrives, right at the end of his expansion sequence, it is not just another example. He shades two faces and gives two perpendicular lengths, so there are two valid expressions for the volume. At this point, students are in a much better place to appreciate just why the cuboid is a special case

The lesson here is straightforward, but counter-intuitive: whenever you introduce a rule, reach for the version that shows the rule cleanly, not the special case that hides it or comes with its own exceptions, even though the special case may appear simpler. Save the special case for the end, when students can appreciate exactly what makes it different.

2. Do not make them do the arithmetic (yet)

Look at the two versions side by side. The original works all the way through to 60 cm³, units and all. Kris stops at V = (5)(10).

The idea you are teaching is that volume is an area multiplied by a perpendicular length. Working out that 12 × 5 = 60 is a separate skill. If a student selects the right area and the right length but slips on the times-table, they get the answer wrong and conclude “I can’t do this”, when they had in fact nailed the thing you were actually teaching.

I pushed Kris on this. Could you not just use easy numbers? His answer was a three-step approach to arithmetic. First, no arithmetic at all: just build the expression. It is far quicker, so you get three or four times as much practice on the concept itself, and you remove the “I got it wrong” spiral for the students whose arithmetic is shaky. Then, once the concept is secure, easy numbers such as threes, fives and tens, to check they can carry the arithmetic inside a longer routine without tripping. Then, and only then, awkward numbers, because the exam will not be kind, with that arithmetic fluency built somewhere else, in dedicated practice, not here.

His line for it was that you do not learn to ride a bike in the middle of the Tour de France. Of course, we want students fluent in arithmetic. Just not while they are also trying to work out what volume is.

There is a diagnosis point buried in here too. If a student gets the answer wrong and there is arithmetic in the mix, it is more difficult to tell whether they have misunderstood the concept or just fumbled a times-table. Take the arithmetic out, and a wrong answer tells you something clean.

3. Do not give them exactly the numbers they need

Nearly every volume of a prism resource you find online does the same thing: gives you precisely the numbers you need, and nothing more.

So does the original sequence. Area is twelve, length is five, multiply them together, next one, change the numbers, repeat.

The problem is that if every diagram hands you exactly two numbers, the task is no longer “work out the volume”. It is “find the two numbers and multiply them”. Students are no longer thinking about the concept of volume.

What you actually want them to do is inspect the diagram, decide which measurement is the area and which is the perpendicular length, and select those. So Kris presents more than one measurement each time. Sometimes the area is given twice, sometimes a length:

Later in the sequence, a length is included that plays no part in the volume calculation. And the orientation of the solid keeps changing, so the length is never just “the number in the same corner of the slide as last time”.

Again, there is a diagnostic gift here. If a student writes 17 × 17 for the first example, you know immediately they cannot yet pick out which measurement is the length. If they write 17 × 17 × 3, you know they are just multiplying all the numbers they can see. Questions where every number given is needed, no more, no less, would never have shown you that.

Here is the lesson: any time your practice gives students precisely the inputs they need, in the order they need them, ask whether they are doing the thinking or just operating a machine. Redundant information and distractors are not there to catch students out. They are what make the task mathematical, and what focuses students’ attention on what matters

One more thing to notice: Kris does all of this, new shape, redundant information, distractors, across just two teacher examples before handing over to the class. I told him I would have needed four… more on this at the end!

4. Take the rule off the screen

In the original, “Volume of a prism = Area of cross section × length” sits at the top of every single example slide.

Kris makes two changes. First, he simplifies the rule to “Volume is area times length”. That is easier to hold in your head than “area of the cross-section times length”. He drops “cross-section” deliberately because it is a term that carries no meaning for a student until you have taught it, and stopping to teach it first creates a barrier to what you actually want them to learn. He also drops “prism”, so that curved solids like a cylinder are not quietly excluded.

Second, he puts the rule on that title slide only, and then takes it away. If the rule is on the board every time they work, students can lean on it every time, and they may never actually commit it to memory. Kris wants it retrieved, not read. So from then on, it is choral response and cold call. What is volume? Area times length. What is the area? Five. What is the length? Ten. They have to produce it themselves.

The general point is simple. Anything you leave permanently on display, students will use as a crutch rather than remember. If you want it in their heads, state it clearly, make it memorable, then take it away and make them retrieve it. And notice that precision was not the goal at the point of introduction. “Area of the cross-section” is more precise, but “area times length” is more memorable, and you can add the precise language back later, once the concept is secure.

5. Normalise the scary numbers

Once the arithmetic is gone, something lovely becomes possible. In the expansion part of Kris’s sequence, the numbers turn deliberately intimidating. A cross-section of √20 cm². A length of 6x. Faces of 27/5 cm² and 5y.

Fractions, surds, algebraic expressions, big numbers: students tend to fall apart at the sight of them. But the fear comes almost entirely from the computation they imply. Take the computation away, and a surd is just a label. The student’s job is still only to pick the area, pick the perpendicular length, and write the expression, which they can now do. So you get to put genuinely scary-looking numbers in front of a class and hand them a high success rate. You are normalising the very thing that usually makes them freeze. Kids feel super smart.

Expansion sequences do carry a risk, because you are showing students something you have not modelled for them. Kris frames it for the class in a way I have since stolen: I am going to show you something that might look a bit weird, and if you get it wrong, that is my fault, not yours. That framing is what makes the risk safe to take.

Beyond volume, the point holds: the intimidation students feel around hard-looking maths is often about the arithmetic bolted onto it, not the concept underneath. Separate the two, and you can build comfort with the notation long before you demand fluent computation with it.

Build it lean, then let the kids show you

I’ll end with a mistake that I would still make despite over ten years of reading about sequence design…

Kris’s teacher-led part is just two examples, and then he hands it to the students.

When I looked at the jump from his first example to his second, I felt it was too big - the cross-sectional face change, the area of that face changes, the nature of the redundant information given changes. I would have used four examples. I would’ve presented the first shape with just one 10cm labelled. Then I would’ve presented that exact same shape again, but with an additional 10cm labelled to draw students’ attention to the fact that they only need one of those 10s. Then I would have changed the shape, and then I would have changed the nature of the redundant information.

Kris described this as over-atomising. In other words, breaking things into more steps than necessary.

The argument for two examples over four is efficiency. If two works, four is wasted time you could have spent on something more interesting. But the deeper point is about how you would ever know. If you build the heavily-atomised four-example version and the students succeed, you have learned nothing about whether the leaner version would also have worked. You have just spent longer. Build the lean version first, and if they succeed, then wonderful! And if they fail, that is fine too, because now their responses tell you exactly where to break it down further.

The catch is that you cannot do any of this from the armchair. Kris pointed out that what made the Direct Instruction programmes he draws on so much for his inspiration so effective was not that they were intellectualised into perfection at a desk. It was that they were field-tested relentlessly, put in front of real children, and revised on the evidence. You cannot, as Kris put it, intellectualise your way to a perfect sequence from an ivory tower.

Design the leanest sequence you genuinely think will work, put it in front of students, and let their responses decide where it needs more steps. Not the other way round.

Over to you

So, some questions to leave you with.

1. Which of the five principles are already embedded in your practice?

2. Which of the five principles are not embedded in your practice, and which do you feel should be?

3. Could you choose one of the principles and build it into your very next teaching sequence?

You can listen to or watch the full episode on the Mr Barton Maths Podcast. Kris wrote up his own account of all ten differences on his Substack, which goes into more detail on each one. And if you like this sort of thing, the community he runs there is full of teachers posting their own sequences and getting genuinely useful feedback on them.

And if, heaven forbid, you’d like me to do more episodes like this, let me know in the comments.

Thanks so much for reading, and have a great week.

Craig

🏃🏻‍♂️Before you go, have you… 🏃🏻‍♂️

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

And checked out Ultimate Retrieval Tool page on Mr Barton Maths?

Here are three blog posts I found interesting this week:

1. Why I hate ‘beautiful books’ by David Didau

Three years ago, I wrote a post called “The Myth of Copying Things Down”. In that post, I made the point that neat books are a poor proxy for understanding. David articulates this point, and much more, beautifully. David argues that the moment leaders signal what they want to see, teachers rationally perform to it. Neat pages can conceal shallow thinking; scruffy ones can hold hard graft. What we’re really rewarding, he suggests, might not be what we think.

2. How Henri Picciotto Designs Great Activities by Michael Pershan

Anyone who spends time thinking about task design will love this closer look at what makes a resource actually work in a classroom. Michael argues that the quiet genius of Henri’s activities lies in three design choices — a single set of instructions that still generates rich mathematics, a stance poised between open exploration and explicit teaching, and a practising teacher’s eye for what genuinely works with real students. The examples are where it clicks.

3. 5 Maths Gems #200 by Jo Morgan

A milestone edition from one of the most reliably useful maths blogs going — 200 gems posts is a staggering run, and this one’s as generous as ever. Jo rounds up her latest haul for maths teachers — self-testing key-facts sheets, a slick new graphing tool, the finally-complete NCETM Key Stage 4 mastery materials, and a couple of self-marking worksheet sites worth a look. It’s a browse-and-grab post, and there’s plenty in here worth stealing.

Have a great weekend!
Craig

🏃🏻‍♂️Before you go, have you… 🏃🏻‍♂️

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

And checked out my Ultimate Retrieval Tool page on my Mr Barton Maths website

Zach Groshell taught for 14 years, mostly fourth grade, mostly in struggling schools - the kind that cannot hold on to staff, teaching the kids who do not arrive pre-loaded with the background knowledge to get themselves through a lesson. He wrote a book called Just Tell Them: The Power of Explanations and Explicit Teaching, which is exactly the exasperated sigh it sounds like. He hosts the Progressively Incorrect podcast. And he is steeped in Siegfried Engelmann’s Direct Instruction - capital DI, the most precise and least fashionable corner of explicit teaching.

He is also, as of about two years ago, the first learning scientist on the team designing the maths apps at Alpha School.

That last detail is why I wanted him on. Zach is the first guest who actually builds the thing - who sits there atomising column addition and deciding what a seven-year-old sees on the screen at four minutes past nine on a Tuesday morning.

He is also, as we will get to, not sending his own daughter to the flagship Alpha school. And that fact tells you what these apps are and what they are not.

Let’s dive in…

1. AI only gives back what your own knowledge lets you put in

Zach uses these models constantly - a second brain, he calls it - and he uses them the way you would hope a learning scientist would: for retrieval practice, for spacing his own study, for worked examples, and then forcing the model to fire probing questions back at him so he is doing the hard thinking rather than nodding along to something he does not really understand.

But the example that landed for me was about building a lesson. Zach could ask Claude to teach adding 4 + 3, take what it gives, ask for it “a different way”, and deploy the result. Instead, he does the opposite. He digs into the actual Engelmann programs, reads, copies and edits the material by hand, and only once he has it right in his head does he hand the assembly to the AI - the genuinely unwieldy part, where one idea has to run across multiple parallel strands with supports fading at the right rate over days. He does the thinking. The model does the typing-out.

So I asked him the question I keep coming back to in this series. Could someone without his hard-won grounding take the same prompts and get the same value? His answer was flat: no. And he has watched it proven at Alpha. The early assumption was that the technology people could generate the lessons themselves - they had the tools, they could make it slick and bug-free. It failed straight away. You need the domain experts in the loop. Engelmann knew this decades ago, which is why he paired his instructional-design brain with subject-matter experts in history, science and maths.

This is exactly the line Dylan Wiliam drew last episode between cognitive offloading and cognitive outsourcing. Zach offloads the assembly; he does not outsource the thinking. And the only reason he can tell the difference - the only reason the model is rocket fuel for him rather than a confident source of plausible nonsense - is the years he put into the source material first.

So here is a question for you. The next time you reach for AI to make a resource, are you offloading the donkey work, or quietly outsourcing the bit that was actually yours to know?

2. The Alpha maths app is built on Engelmann, not a chatbot

If your mental image of an Alpha maths lesson is a child chatting to an AI tutor, drop it. There is no chatbot.

What a child actually meets is a largely bare screen. The maths populates it slowly - dots, lines, a column building up - with slight animations, many of them hand-made because the AI was not good enough at it, and no on-screen teacher narrating over the top. Talk while you change something on screen, Zach points out, and you can interfere with the very thing the child is trying to see. The child rarely watches for more than about 9 seconds before they have to respond. The steps are tiny. Several strands run in parallel and get chained together, typically across three exposures over three days, before a timed mastery test. Roughly 90% of any lesson is review; only about 10% is the new, slightly harder thing. As Zach put it, it is cognitive load theory in a bottle.

This is capital-DI translated to a screen, and it speaks directly to the disagreement Kris Boulton left us with in episode ten. Kris’s view was that AI has cracked the practice - it will generate endless well-formed task sets - but not the teaching, the carefully sequenced initial instruction. Zach mostly agrees. The sequences are still designed and field-tested by humans, then hand-animated; he was clear that the AI cannot yet generate the categorical teaching sequences itself. So the honest framing is not “the AI teaches maths”. It is “humans designed the DI instruction, and the app delivers it with a patience no tired teacher at four o’clock can match.” The teaching is human. The delivery is the machine.

3. “Just tell them”: carefully chosen examples beat long explanations

Perhaps a more apt title for Zach’s book would have been “Just Show Them” because he warns against lengthy explanations and instead favours carefully sequenced examples and non-examples. This is true whether instruction is delivered by a teacher or an app

A bad teaching sequence leans on meandering explanations and then dumps a pile of practice problems on the child. That is discovery learning in the guise of a worksheet. Good DI does the reverse. It rules out wrong interpretations using examples and non-examples, often as a plain yes/no: this is correctly regrouped, this is not correctly regrouped - which of these is correctly regrouped? You build the conceptual and procedural pieces this way, chain them, and the full routine clicks the first time. The child gets it, and feels clever for getting it. These are the same categorical atoms I wrote about here.

Regular readers will hear the echo of Daisy Christodoulou from the very first episode, who argued that explanations are massively overrated - they overwhelm the child, they hand you no data on whether anything has landed, and they are wildly inefficient next to a well-chosen sequence of examples. It is also, more or less, the case for the Silent Teacher approach I have banged on about for years: show, don’t narrate, then check.

So a question worth sitting with. Think about the last genuinely new idea you taught. Did you explain it - or did you choose a sequence of examples and non-examples so tight that the wrong interpretation never got off the ground?

4. Good retrieval retires itself

This was the idea I had not heard put quite this way before, and I think it is a good one.

A lot of retrieval, Zach argues - especially in maths - is what he bluntly calls dumb retrieval. We pick a small skill, decide it is “a thing”, and dutifully space it out across the whole year on an ever-expanding interval. But if a child is now doing column addition, identifying the digits and setting out the problem, is already baked into the task. Re-testing those skills in isolation is a waste of their time. The skills have been subsumed, so they should be retired, and what you retrieve instead is the bigger routine - at increasing levels of sophistication, with more steps and more context as the child’s knowledge matures.

The classroom takeaway is sitting in your Do Now right now. Have a look at it. How many of those questions are retrieving atoms your students mastered and subsumed weeks ago? Retire them. Retrieve the routine instead.

5. He builds these apps - and he did not send his daughter to Alpha

I love a twist, and Zach produced a good one at the end of our chat.

Zach separates two things that usually get bundled together: the app itself, and what a child does with the rest of the day. “Alpha” is just the name for the schools at the expensive end of this network - apps in the morning, life-skills and projects in the afternoon. But the same apps are being dropped into very different models. There is Texas Sports Academy, for the child who currently suffers through a school that cannot keep order and then comes alive at half three for sport - two hours of apps in the morning, sport in the afternoon, and crucially a model cheap enough to be heading for the US voucher system. There is Montessorium in Austin, where the mornings are unapologetic Engelmann DI and the afternoons are human-led, small-group, Montessori-style academic work.

Zach chose Montessorium for his daughter. Not because he doubts the apps - he builds them - but because she loves reading whole books, journaling and drawing, and the sport-in-the-afternoon model would do nothing for her. The apps, in other words, are not a school. They are an efficient engine for getting foundational skills into a child’s head in about two hours, and that engine can live inside whatever model fits a family’s values. The morning is the foundational skills. The afternoon is where the values actually live.

This reframes the question I have been chewing on since Becky Allen called Alpha a benign laboratory. The interesting thing was never going to be the $40,000 flagship. It is whether the engine can be lifted out and dropped somewhere it reaches the kids who need it most. And on the worry that haunts this whole topic - Dan Meyer’s line that no child wants to impress an app - Zach’s answer is that they still want to impress the guide, that the daily launch ritual is doing real social work, and that an app pitched permanently at the child’s level removes the usual reason to act up. The child who sabotages a lesson, he argues, is almost always the child who is lost and demotivated by failure. Give them something they can actually master, and the motivation to misbehave quietly drains away.

What stayed with me was where Zach put his hope. Not the flagship campuses, but in the rural kid with no physics teacher, the kid whose dad knows something is wrong but cannot afford Kumon, the kid who could rattle through a year of algebra in three months and is currently being made to wait. If this ends up as a parallel option for the children the system already fails, that is an outcome worth taking seriously.

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 Zach’s book, Just Tell Them, and his podcast, Progressively Incorrect, which is one of the best places going for the actual moves of teaching.

This is the twelfth 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

6. Dan Meyer: podcast and write-up

7. Bibi Groot: podcast and write-up

8. Becky Allen: podcast and write-up

9. Gráinne Hallahan: podcast and write-up

10. Kris Boulton: podcast and write-up

11. Dylan Wiliam: podcast and write-up

Twelve episodes in, and now that we have finally seen inside the box, I would love to know where you land on Alpha. Does seeing the Engelmann engine under the bonnet make you more interested or less? Does the “just tell them” case change how you would teach your next new idea? How much of your Do Now is retrieving things your students retired weeks ago?

Thanks so much for reading, and have a great week.

Craig

🏃🏻‍♂️Before you go, have you… 🏃🏻‍♂️

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

And checked out Ultimate Retrieval Tool page on Mr Barton Maths?

Here are three blog posts I found interesting this week:

1. Do Androids Dream of Electric Seating Plans? by Joel Kenyon

I love AI, and I’m genuinely excited about what it could do for teaching… which is exactly why a dystopian vision of it all going wrong has stuck with me all week. Joel imagines education in 2045: AI deepfake teachers trained on real staff, classes of 250, and human teachers reduced to minimum-wage behaviour monitors. Joel argues that when cost-cutting drives edtech, the relationship at the heart of teaching is the first casualty.

2. Scaffolding is a Trap by Dominic Salles

I also love a bit of “I Do, We Do”, so I was lured in by Dominic’s latest post. He argues that scaffolding — unless you plan its removal from the start — becomes the enemy of hard thinking. Tasks should be simple, he says, but thinking should be complex from minute one. He even reckons we’ve misread Rosenshine’s 80% all along.

3. Wait Time by Jonathan Nicolas Jarrett-Kerr

Jonathan makes the case for wait time — think time — as one of the most inclusive things a teacher can do, precisely because it’s so hard to hold your nerve. He argues it reaches well beyond questioning — into whiteboards and quizzes too. The tricky bit is judging how long.

Have a great weekend!
Craig

🏃🏻‍♂️Before you go, have you… 🏃🏻‍♂️

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

And checked out my Ultimate Retrieval Tool page on my Mr Barton Maths website

In September, Dylan Wiliam will have spent 50 years in education. He started out as a teacher, had an argument with his head teacher, and was offered a research fellowship the next day. From there it was lecturer, professor, head of department, assistant principal of King’s College, a spell at the Educational Testing Service in the US, and Deputy Director of the Institute of Education. For the last 15 years, he has been a freelance, self-financing academic, doing the work he thinks is important and funding water wells in Africa with the proceeds of his workshops.

For many of the teachers reading this, Dylan is the reason they care about formative assessment at all. So I was thrilled to get him back on the podcast for this series.

And here is the thing that made this conversation different from the other ten. Every guest so far has, in their own way, planted a flag. Adam Boxer told me AI will not revolutionise education because nothing ever does. Carl Hendrick told me this time is different. Most of my other guests landed somewhere they could defend.

Dylan refused to plant a flag. He told me he agrees with Carl that this time is different. Then, in almost the same breath, he told me he agrees with Adam too. When I pointed out that these are opposite conclusions, he just smiled and said yes, they are, and he can’t choose, and isn’t that interesting?

In a series full of confident people - and after my recent dabblings in the world of LinkedIn, where I have never met a more certain community in my life - there was something genuinely refreshing about a man with 50 years of expertise telling me, repeatedly, that he was less sure than he used to be.

Let’s find out why…

1. Dylan thinks this time is different… but he is also cautious

Dylan was sceptical when ChatGPT launched. He has lived through enough false dawns to have earned his caution. So why does he think this time things are different?

His reasoning is simple - it is just the sheer number of things we confidently said AI would never be able to do, that it now does. AI can’t be creative… except the Boston Consulting Group found it generates better product ideas than people. AI can’t be empathetic… except, Dylan says, it already comes across as more empathetic than the typical doctor. As he put it: “I’ve stopped saying AI will never be able to do X because I’ve said that so many times and been wrong.”

However, he is deeply alert to one thing the optimists tend to skate over: the feedback loops in education are agonisingly long. AlphaGo learned Go by playing itself a million times in a few days. You cannot do that with a child’s education. Dylan reached for the Reading Recovery studies in the US: randomised controlled trials showed real gains by the end of first grade, and then, three years later, those same children were about six months behind. Things that look wonderful in the short term can turn out to lay bad foundations. And when we’re talking about a child’s future, that is a big risk to take.

2. Offloading is not outsourcing

Dylan draws a careful line between two things that often get muddled. Cognitive offloading is when you jot down the working for a long multiplication. Those marks on the paper extend what Merlin Donald calls the external memory field - they become part of your working memory and make your thinking more powerful. You are still doing the thinking. The tool is boosting it.

Cognitive outsourcing is when you simply hand the whole job over and let the machine do the thinking for you.

Regular readers will recognise both ideas. I have written before about Paul Kirschner’s framing of offloading in my piece on resource creation, and about what Peps Mccrea calls cognitive outsourcing in Is Maths Homework Dead? Dylan’s contribution is the test he applies, and it is beautifully simple. Before he uses AI for anything, he asks himself one question: Is this something I want to get better at?

If the answer is no, he outsources happily. He is not interested in becoming a better writer of condolence letters, so he is glad of the help. But if it is something he wants to improve at - data analysis, say - he does the work himself and only uses AI as a final check at the end to ask: “Have I missed anything?”

Now apply that question to a fourteen-year-old with a Sparx login and ChatGPT open in the next tab. Do they want to get better at solving equations? Almost certainly not, in the moment. So they outsource, and the thinking that was the entire point of the homework never happens. This is exactly what Bibi Groot pointed out in episode seven - the Turkish students who practised with an unconstrained chatbot did worse on the no-AI exam, because they had outsourced the cognitive work and had less in their heads when it was taken away.

The offloading/outsourcing test gives us a clean way to talk to students about this. Not “AI is banned” and not “AI is fine”, but a genuine question they can learn to ask themselves: Is this a thing I am trying to get good at? Because if it is, letting the machine do it is robbing your future self.

Here are two questions for you. Where in your own practice are you outsourcing something you’d actually be better off getting good at? And how would you teach a class of Year 9s to tell the difference?

3. AI’s biggest promise in formative assessment is turning the average teacher into a more experienced one

I love nothing more than a bit of formative assessment chat, so I loved this section of our conversation. Dylan sees three distinct roles for AI in formative assessment: a proactive role (before the lesson), an interactive role (in the moment), and a retroactive role (afterwards).

The proactive one is where he is most convinced. If you have taught a topic a dozen times, you already know the wrong turns students take. But if you have never taught it before, you don’t know what you don’t know. Have AI surface the common misconceptions for, say, the causes of the First World War, and you can design teaching that heads them off in advance. As Dylan put it, this “makes a novice teacher into a slightly more experienced teacher.”

The interactive role is more speculative but genuinely exciting: hints arriving on a teacher’s iPad in real time, or AI suggesting how to group students on the fly based on a quick survey question - homogeneous groups so you can pull the strugglers together, or deliberately mixed groups, which we know (from Mevorach and Kramarsky’s work) can be powerful for metacognition if they’re carefully structured. The teacher stays in charge and ignores the suggestion whenever they know better.

But Dylan put a stake in the ground here that I think is important, and it leads straight into the next takeaway. When I described our DQR pilot - the QR-code system we’re running in primary schools that captures which child chose which answer, flags the likely misconception, and offers a follow-up question - he was supportive of the follow-up idea (he even suggested going further, with three follow-ups testing increasing degrees of transfer, so you find out how generalisable a student’s understanding really is). But he was wary of the data capture itself…

4. We have too much data and not enough evidence

For all that Dylan has spent his career thinking about assessment, the most striking thing he said was a warning against collecting more data.

He told me a story from his master’s degree. He wrote a 30-question diagnostic test for his Year 8 class, got the data back question by question, and it hit him: from a 20-minute test, he had more data than he could act on in an entire term. The bottleneck was never the data. It was turning data into something he could actually use.

He has a lovely phrase for this, borrowed from the intelligence analyst David Schum: data becomes evidence when it is used in support of a conclusion. Most of what we collect in classrooms is data. Very little of it ever becomes evidence. And so Dylan’s instinct runs against the spirit of the age. He does not believe in data-driven decision-making. He believes in decision-driven data collection - work out the decision you are trying to make first, then collect only the data that will inform it.

This is where he is a useful counterweight to Carl Hendrick’s prediction from episode four. Carl thinks the enormous datasets coming out of platforms like Alpha School will finally reveal the things about spacing and interleaving that we have never been able to pin down - the optimal spacing schedule, for example. Dylan is not so sure. His worry is that this is found data - collected because it happened to be lying around, not generated to answer a specific question. In 35 years of research, he says, found data sets have almost never given him the data he actually needed. Simon Woodhead made a related point in episode five about how hard it is to turn a decade of student responses into something genuinely predictive.

And then there is the ethics of it. If we want classrooms where children feel safe making mistakes, Dylan asks, why on earth would we record every single one of them? He likes the ephemeral nature of a mini-whiteboard, of finger-voting, of a card held up and then lowered. The mistake vanishes. Start permanently logging every child’s every wrong answer and, in his words, “it looks like we’re going towards 1984.”

The entire series has had a quiet assumption running through it that more data is better. Dylan’s challenge is the right one. Before you capture it, ask what decision it serves and what it costs the child to have it recorded. We may already have too much data and not nearly enough evidence.

5. The question that matters most isn’t how AI changes teaching - it’s what we’re teaching for

When I asked Dylan the question I ask everyone - what should I have asked that I didn’t? - he didn’t reach for a clever AI use case. He went somewhere bigger.

Dylan argues that so much of what we do in upper secondary school is precisely the kind of thing AI now does extremely well. If the workplace these students are heading into has been reshaped by AI, then we may be carefully preparing them for a world that no longer exists. As he put it: “I think we might be preparing students for the wrong kind of future.”

He is not saying we should stop teaching knowledge. Quite the opposite. He pointed straight back to Barbara Oakley’s episode - you need facts to think with, and you don’t even know what questions to ask without a solid grounding. The issue isn’t whether knowledge matters. It’s what else an education is for once a machine can do the things currently taught.

Two ideas of his have stayed with me. The first is Baumol’s cost disease - the observation that a Mozart string quartet takes exactly as many musicians and exactly as long to play as it did 200 years ago, because some things simply don’t yield productivity gains the way construction does. Dylan’s hunch is that good teaching is more like the string quartet than the building site. If he’s right, the dream of doing more with fewer teachers runs into a hard wall, and the real threat is that we index teacher pay to prices rather than wages and slowly make the job impossible to recruit for. Dan Meyer made a similar argument in episode six - that the honest solutions to our problems, paying teachers more and shrinking class sizes, are political, not technological, and get ruled out before the conversation even starts.

The second is a vision of how schools might gradually shift. Dylan sketched a spectrum from the rich, exploratory environment of the early years to the near-total independence of his PhD students, who he’d meet once a fortnight with a list of things to go away and figure out. His prediction, ten years out: Key Stage 3 will look much as it does now, Key Stage 4 will have more independent work, and sixth form could look radically different - far closer to that PhD model of independent study. The thing most likely to stop it? Schools so worried about A-level grades that they spoon-feed students into good results and then send them off to university unable to work on their own.

Dylan isn’t a wholesale fan of the creativity gospel, but he holds onto the idea that one of the jobs of a school is to show children things they’d never otherwise meet, so they can find their thing. If AI really does take on more of the work, and if more of life ends up as leisure, then what we squeeze out of the timetable first - art, music, drama, the chance to improvise in a jazz band or do improv comedy - might turn out to be exactly the stuff that matters most.

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. Do check out Dylan’s YouTube channel, where he shares findings as he comes across them. And the water wells charity he mentioned at the top of the episode is well worth a look.

This is the eleventh 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

6. Dan Meyer: podcast and write-up

7. Bibi Groot: podcast and write-up

8. Becky Allen: podcast and write-up

9. Gráinne Hallahan: podcast and write-up

10. Kris Boulton: podcast and write-up

Eleven episodes in, I’d love to know where Dylan’s comfortable uncertainty leaves you. Does holding two opposite views at once feel like wisdom or like a cop-out? And on the question that’s stuck with me most: if AI really can do the text manipulation, what do you think upper secondary school is now for?

Thanks so much for reading, and have a great week.

Craig

🏃🏻‍♂️Before you go, have you… 🏃🏻‍♂️

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

And checked out Ultimate Retrieval Tool page on Mr Barton Maths?

Here are three blog posts I found interesting this week:

1. Why principles beat practices every time by Mark Enser

The gap between doing something in the classroom and understanding why you’re doing it is often large and significant. Mark argues that CPD prescribing practices without their underlying principles breeds “cargo cult teaching” — perfect rituals, absent understanding. Practices are routes; principles are maps. And a profession trained only to follow routes, he warns, slowly loses the people who could ever draw a new map.

2. Managing Call-Outs in the Classroom by Jonathan Jarrett-Kerr

Every teacher knows the eager hand that answers before you’ve finished asking. Jonathan argues that praising a call-out — “well done, but don’t shout out next time” — backfires, because the dopamine of being right lands before the correction and quietly rewards the behaviour. The smarter move is to get upstream of it — by reordering a single question.

3. Beyond the ‘To-Do List’ Curriculum: Why School Mathematics Needs to Focus on Big Mathematical Ideas by Colin Foster

And to finish, a treat for my fellow maths teachers — and quiet vindication for anyone who’s ever suspected that the many topics a student struggles with are really one or two things in disguise. Colin argues that school maths too often becomes a fragmented “to-do list” of procedures — rehearsed, then forgotten. His alternative organises the whole curriculum around just five “Big Ideas” that cut across topics, starting with thinking multiplicatively. Many topic-specific struggles, he suggests, are really one shared weakness wearing different masks.

Have a great weekend!
Craig

🏃🏻‍♂️Before you go, have you… 🏃🏻‍♂️

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

And checked out my Ultimate Retrieval Tool page on my Mr Barton Maths website

Here are three blog posts I found interesting this week:

1. Why personalising education based on student interests doesn’t work by Daisy Christodoulou

The promise that AI will finally “personalise” learning is everywhere right now, so I loved reading a proper, evidence-grounded pushback. Christodoulou argues that personalising lessons around student interests — teaching equations “through the medium of Taylor Swift” — fundamentally misunderstands how analogies work. The link to prior knowledge has to fit the content, not the learner’s hobby.

2. Why Knowledge-Rich Curriculum Never Wins by Robert Pondiscio

It’s written about American schools, but every one of these obstacles will feel familiar to anyone who’s championed a knowledge-rich curriculum over here. Pondiscio argues that the evidence war over knowledge-rich curriculum is won — cognitive science, the science of reading, even AI all confirm it. So why does it keep losing in practice? Robert shares 19 reasons!

3. Lessons Worth Arguing About by Becky Allen

One for my fellow maths geeks to finish. Allen argues that teachers keep arguing about how to introduce negative numbers or fractions not because the profession is muddled, but because these first-encounter lessons are where sound instructional principles collide.

Have a great weekend!

Craig

🏃🏻‍♂️Before you go, have you… 🏃🏻‍♂️

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

And checked out my Ultimate Retrieval Tool page on my Mr Barton Maths website

Here are three blog posts I found interesting this week:

1. In Defense of Doing Things Together by James Tucker

One theme I keep returning to in my AI in Education interview series is the question of whether learning is inherently social. Tucker argues that machine-driven learning is seductive precisely because it works — and that’s the danger. Once efficiency rules, the messy business of learning the same thing together starts to look like waste. What gets quietly optimised away, he warns, is something schools exist to protect.

2. The Idiot Who Talked About Discovery Learning at Interview by Chris Reid

Having travelled my own road from well-meaning chaos to explicit instruction, I’m always drawn to writing brave enough to complicate the tidy version of that story — and Chris does it superbly. Reid argues that the real risk isn’t choosing direct instruction over enquiry — it’s a quieter mutation he calls “total instruction,” where teacher judgement gets stripped out entirely. The evidence on method, he shows, is genuinely mixed. The question we should be asking is a different one altogether.

3. The Use of Context in Assessment by Matthew Evans

Matt Evans is a deep thinker, and this post is a great example of that. Evans argues that we fixate on the wrong thing with “authentic” assessment — whether a task mirrors the real discipline. A borrowed idea, semantic gravity, offers a sharper lens: what counts is how a good answer moves between the concrete and the abstract.

Have a great weekend!

Craig

🏃🏻‍♂️Before you go, have you… 🏃🏻‍♂️

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

And checked out my Ultimate Retrieval Tool page on my Mr Barton Maths website

Kris Boulton once told me you should never ask a student a question they do not already know the answer to. He has spent the best part of 15 years building maths teaching that makes that possible — sequences of examples and non-examples so carefully ordered that failure is, in his words, designed out. He now runs Unstoppable Learning with Naveen Rizvi, working with schools from Australia to Texas.

So when Kris came on the podcast and told me he is now having more conversations with AI than with real people, and that he had just spent a working day inside Alpha School in New York, I knew this one was going to be good.

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

1. Alpha School has no teachers — and that is the whole point

Before I spoke to Kris, I had heard Alpha described as a school where the teachers are called “guides.” Kris stopped me right there. They do not call their teachers guides. They have no teachers. The guides are adult supports — coaches and mentors — and they are not permitted to help a child with an academic question.

Why on earth not? Kris quoted Joe Liemandt, the principal of Alpha School: they want the cleanest data set they possibly can get. If a child is not learning, they need to know that it is because the app is not good enough. If a guide quietly steps in and fixes a problem, that signal is lost. So the bet is enormous — that the app can be made good enough on its own, with the human role stripped right back to motivation.

Kris suggested that this approach might be more appealing to primary teachers — ask a primary teacher what they teach, and many will say “I teach children, not subjects.” But secondary teachers tend to love their subject. We love sharing it with someone meeting it for the first time. At Alpha, you do not get to do that at all. As Kris put it, you do not get to share any of the things you love and know about mathematics with anyone.

2. “Rich parents are rich parents. Rich kids are just kids.”

The standard objection to Alpha is the one I have made myself: How much can we learn from a school that charges top-of-market fees and selects for wealthy, motivated families?

Kris argued that whilst rich parents are rich parents, rich kids are just kids, with everything that comes with being a child.

He saw it first-hand. One lad, around five or six, was not having any of it — dysregulated, struggling, not engaging. The wealth of the parents did not magic that away.

So why does this matter? Because the easy dismissal — “it only works because the kids are loaded” — does not survive contact with the place. The harder questions remain. Does it work for children whose parents are not bought in? Does it work without the fees? But Becky’s point from her episode holds: you answer “Can this work at all?” before you answer “Can this work for everyone?”

3. They pay the kids — and the thing that bothers us is not the thing that should

Here is the Alpha detail nobody else in the series mentioned… the kids get paid. Hit your daily minimums, and you earn somewhere between one and two dollars a day. For the little ones, it is “Alpha bucks” for a curated store. For the older ones, it goes onto a debit card they can spend anywhere in New York.

I can feel some of you bristling. I bristled too. But Kris made me think harder about why.

The usual objection is that extrinsic rewards crush intrinsic motivation — pay a kid to do maths, and they will only ever do maths for money. Kris thinks the research and his own experience do not bear that out. Get genuinely interested in something, and you keep wanting to know more, paid or not.

So what is the discomfort? Kris’s sharper read is that paying kids pokes at a buried sense of gratitude. Thousands of pounds are already being spent on a child’s education, for their benefit — and now we are paying them a few dollars on top just to turn up and accept it. That feels off. But then he turned it on us: how many teachers grumble about attending CPD that is entirely for their own benefit? The delayed, abstract reward of “this will make you better” is, it turns out, just as hard for adults as for kids.

Whatever your moral position, Kris’s point is the one you cannot argue with: as a way of motivating people to show up and do the work, it is effective. For more on motivation and rewards, check out one of my favourite EconTalk episodes with Roland Fryer.

So, there are three additional points about Alpha School to add to those that we’ve heard throughout this series. Carl Hendrick was generous about the Alpha model. Becky Allen called it a benign laboratory worth building in. Adam Boxer said it could lead to the death of expert teachers. Where do you sit on that spectrum?

4. AI has cracked the practice — but not the teaching

This is where the conversation turned from a bit geeky, to full-on geek mode… and I loved it. It’s time to dive into the wonderful world of Atomisation, and how it relates to AI.

As long-time listeners to my podcast will know, Kris splits the learning experience into stages: initial instruction, initial testing, expansion (applying the idea to ever more novel cases), and then a long tail of practice over days and weeks. And his verdict on where AI sits is precise.

For the practice — the task sets that come after teaching — AI is there. It is perfect, in his words. Set the constraints, tell it every question must match the form of what was taught, and it will generate thousands of usable tasks in hours instead of the weeks it would take by hand. He could not have built his new Unstoppable Arithmetic programme without it.

But for the initial instruction and the expansion — the carefully sequenced examples, non-examples and categorical atoms that are the heart of his work — he has not seen it work yet. He has fed Claude every script for every video on his platform, whole curricula from Australia and the US, and it still gets the type of atom wrong every time. Not always badly. But too imperfect.

Why the split? Because the sequencing barely matters in the practice phase — every question is one the student has already met under tight instruction. It matters enormously in the instruction phase. And that is exactly the part the AI cannot yet feel its way through.

I have a slightly different view. I’ve managed to get Claude to produce teaching sequences for categorical atoms that meet the standard I can design on my own. And I’m able to do it in a fraction of the time and enjoy the experience much more as I collaborate with my super-smart, super-patient AI colleague. I’ve shared my efforts here.

Of course, my standards are likely to be much lower than Kris’s, so my point may be moot!

5. The average teacher is not listening to this podcast — and that should shape the advice

I put a crude scenario to Kris. A maths teacher has a finite amount of time to prepare for a lesson tomorrow. They have three options: build it by hand, download something already made, or open Claude or ChatGPT and generate it. Which leads to the best outcome?

Kris answered the easy half instantly. The average teacher will take what is already made, deliver it, and move on. And — his line, not mine — the average teacher is not listening to this podcast or reading this newsletter. We forget, constantly, what average in the system really looks like. If you are reading this, you are likely to be top tier!

Then he reframed the question. It is never one lesson tomorrow. It is 23 lessons next week. Whatever you choose has to work across all of them. Build by hand for 23 lessons? Out. Wrestle every one of them out of an AI? Out for most people, because of the time and the cost — we both agree that you need the £75-to-£200 models for this quantity and complexity of work, and this is not something most teachers will fund from their own pocket.

So who should play with AI? Well, like anything, what you get from it is a function of how you already think. Some people race through books and learn little. Others read slowly, argue in the margins, and learn loads. AI is the same. If you have the subject knowledge to catch its errors and push back, and the time and patience to do so, it can supercharge you. If you do not, you are the non-specialist putting wrong answers on the board because you could not spot them — exactly the danger Adam Boxer and Gráinne Hallahan warned about earlier in this series.

Here is what I think is a good way to dip your toes into AI resource creation…

Pick one lesson you’re teaching next week. Plan your own resource from scratch. Prompt the AI to do the same. Then compare the results. What does your resource do better? What does the AI do better? Then feed your conclusions back to the AI, share the resource you created, and see if, together with a few more back and forths, you can come up with something that is greater than the sum of its parts. That back-and-forth is where the learning lives. The trap is letting it do the thinking for you.

Boxer versus Hendrick: where does Kris land?

(That’s my attempt at a clickbaity subheading)

Adam Boxer told me AI will not revolutionise education because nothing ever does. Carl Hendrick told me this time is different. So where does Kris sit?

Roughly with Adam, but for a different reason. He thinks the questions everyone keeps asking — will Alpha scale, will it work in a UK comprehensive — are the wrong questions, because nobody knows. What he is sure of is the cost of getting there: tens of billions in capital, retraining, and the unavoidable question of what happens to all the teachers if you flip the model and there are no teachers in the building? Change, if it comes, will come bit by bit, over decades. And then one day everything is different, and you realise it took 30 years.

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 the Unstoppable Learning Substack, where Kris and Naveen post nearly every week, and the Unstoppable Learning website.

This is the tenth 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

6. Dan Meyer: podcast and write-up

7. Bibi Groot: podcast and write-up

8. Becky Allen: podcast and write-up

9. Gráinne Hallahan: podcast and write-up

I would love to know: having seen inside Alpha, are you more tempted by the two-hours-and-the-afternoon-back model, or more convinced it could never work here? And on the AI question — are you using it to learn, or letting it do the thinking for you?

Thanks so much for reading, and have a great week.

Craig

🏃🏻‍♂️Before you go, have you… 🏃🏻‍♂️

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

And checked out Ultimate Retrieval Tool page on Mr Barton Maths?

Here are three blog posts I found interesting this week:

When Prior Knowledge Helps, and When It Doesn’t by Efrat Furst

I’ll read anything Efrat writes, and this is another classic. Furst argues that the famous maxim — learning depends most on what the learner already knows — needs a sharper edit: it’s not just what learners know, but how that knowledge is structured. Whether it’s shallow or deep, she suggests, should be the first thing teachers weigh — and it flips everything that follows.

What is procedural fluency? by Bill McCallum

A lovely example of someone refusing to let a slippery term stay slippery. McCallum argues that procedural fluency has been muddied by loose talk of “flexibility” and “automaticity,” so he rebuilds the term from the ground up. The result is a tighter, more functional definition — one centred on what fluency is actually for.

Slow Reveal Graphs Are Awesome by Dylan Kane

There are few things in life I like more than a maths website I’ve not previously discovered. Kane argues that school maths shows pupils such a narrow slice of what graphs can look like that one of his students thought a perfectly normal graph was “facing the wrong way.” His fix is a simple daily routine, via a cracking website, and a quiet case for letting kids learn surprising things about the world.

Have a great weekend!

Craig

🏃🏻‍♂️Before you go, have you… 🏃🏻‍♂️

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

And checked out my Ultimate Retrieval Tool page on my Mr Barton Maths website

If you find this tool useful, a great way to show your thanks is to spend 30 seconds filling out this survey to help us decide what tool to build for teachers next. The survey has one key question:

Thanks so much, I really appreciate it.

Okay, let’s get two things out of the way straight away:

1. Calling something The Ultimate Retrieval tool sets you up for a massive fall. In the words of my wife, it’s a bit cocky, isn’t it?

2. I am very aware that the last thing the world needs is yet another maths question-generating tool.

So why have I spent my evenings, weekends, and most of half-term building one? Because it does six things that I think matter and, as far as I am aware, that no other tool does. Let me walk you through them, and then show you four ways you might use it.

If you want to jump ahead, you can have a play with it here: The Ultimate Retrieval Tool. It is completely free. No adverts, no pop-ups, no logins, nothing dodgy.

Six things The Ultimate Retrieval Tool does

1. A code that brings your quiz back

Build a quiz outside the lesson, get the topics and questions exactly how you want them, then lock it in. The tool gives you a short code, printed on the worksheet alongside a QR code.

Type that code back in, or scan it, and your quiz reappears — questions, answers, and a set of fully worked solutions. So you can build the quiz at home and project it in class. Or build it in class and let students revisit it at home. Share the code with a colleague, and they get your exact quiz too.

The frustrating thing about most generators is that you build something good and then cannot get it back. A quiz you made at the weekend is no use if you cannot regenerate it on Monday morning. And a quiz you built before the lesson and printed off is not much use if you can’t then access it in the lesson to call up all the answers and worked solutions. With this tool, this is not an issue.

2. Four questions or ten — and it fits on one side of A4

You can build a quiz with four or ten questions. I chose those two numbers on purpose. Four is ideal for a retrieval Do Now. Ten is ideal for a weekly Low-Stakes Quiz.

And because the number is fixed, I can guarantee the printing works. A ten-question quiz fits exactly on one side of A4. A four-question quiz fits two to a side. No awkward half-pages, no font shrinking question by question. Print, photocopy, done.

3. Block one skill or interleave many

Practise any single topic, or any combination of topics. Whatever you have just taught, you can give students blocked practice by picking one or two related skills. Or you can tap into the power of mixed practice — interleaving — by choosing skills from right across the maths curriculum.

Mixing topics forces students to do the thing exams demand and classrooms rarely rehearse: working out which method a question is even asking for. Block practice tells them which skill to use. Interleaving makes them choose.

I’ve got big improvements planned for the interleaving engine, but it works okay for now.

4. Geometry, not just arithmetic

AI can now generate images far better than it could even a few months ago. I wrote about this here. The Ultimate Retrieval Tool renders diagrams, not just lines of text.

Right now, I am working my way through Number. But it will grow to cover angle facts, statistical diagrams, trigonometry, circle theorems — the topics these generators usually leave quietly missing because the diagrams were too hard to generate.

5. A fresh question on the same skill, in one click

Every question has a ↻ button. Press it, and you get a new question on the same skill, with different numbers, a different diagram, a different context.

There are three nice ways you could use this.

1. When you are planning, the first question that pops up might not suit your class, so click to get another.

2. In the lesson, your students struggle with a question, you offer an explanation, you click for a follow-up, which students answer on mini-whiteboards. Now you have evidence about whether your explanation actually landed, instead of a hopeful “Does that make sense?”

3. Share the code with your students, so when they get home, they can generate as many versions of a question as they like, until they have nailed the skill.

6. A whole follow-up quiz, in one click

Once a quiz is locked in, one button builds a follow-up of the entire quiz: the same skills, in the same order, with fresh numbers throughout.

So if you run the same topics across a week of Do Nows, you can generate five versions in seconds. Or you can set a low-stakes quiz on Thursday, let students revise it over the weekend, and hand them a fresh version on Monday. Same skills, new questions, and a clean read on what has actually stuck.

Four ways to use The Ultimate Retrieval Tool

1. The Do Now

Pick four questions from four different topics. I am a fan of mixed-topic Do Nows unrelated to the lesson's content to ensure students don’t forget what they once knew. Put the quiz on the board and have students answer on mini-whiteboards. For any question they struggle with, offer an explanation, then hit ↻ to get a new question. Students answer this follow-up question on their mini whiteboards, which gives you evidence in the moment as to whether your explanation made sense. For more on the do now, check out my book.

2. A weekly Low-Stakes Quiz

Low-stakes quizzes are one of my all-time favourite retrieval opportunities. Print a ten-question quiz, students do it in silence, go through the answers on your board (each question has an answer and a fully worked solution), and students mark it themselves. Then they take their quiz paper home — code and all — and use it as an interactive revision document, accessing answers, full working out, and unlimited question generation. For more on low-stakes quizzes, check out my book.

3. Interleaved practice

You have taught a concept and students seem to have nailed it. The real test is whether they can still answer questions on it when those questions are jumbled in with everything else. Select a ten-question quiz, click the skill you have just taught, throw in a few skills from earlier in the year, and there is your mixed-topic worksheet. The struggle of switching between topics is the point, not a side effect.

4. Homework

You shouldn’t pick favourites, but this might be mine.

I have argued before that homework as a check for understanding is dead. It does not matter whether you set it on paper or on a platform. A student can take a photo, upload it to ChatGPT, copy the working… and score full marks every time. If the purpose of homework is to do it and score well, and that is what they are held accountable for, then the kids who most need to do it are the very ones who will cheat.

The only solution is to flip the purpose of homework. Set a printed ten-question quiz for homework and tell students they can use whatever they like — AI, the tool’s own follow-up questions, a textbook, a revision guide, a friend, you genuinely do not mind. Then, next lesson, hand them a follow-up version on paper (you can generate this with a single click), no technology in the room, and tell them that their performance on this no-tech follow-up will be what they are held accountable for.

Why does this work? Because you have stopped trying to police the homework, which you cannot control, and started assessing the thing you can — the in-class quiz. The homework is no longer the assessment. It is preparation for the assessment. And the moment that becomes true, the incentive to cheat disappears, because the only way to do well on Monday is to have actually learnt something over the weekend.

Why I built The Ultimate Retrieval Tool

There are two reasons why I built this tool and why I’ll never charge for it.

First, I’ll be honest… I am addicted. I love building things with my mate Claude. I do not just say “build me a generator on adding fractions.” I have a proper back-and-forth about which numbers to use, what constraints to set, what variations to build in. I find that stuff fascinating, pushing the technology to its limits, and I am learning a huge amount from it.

Second — and I am very aware of how corny this sounds — I want to give a bit back. I have learnt so much from maths teachers around the country and around the world over the last 22 years. People have been kind enough to find my work useful, whether that is booking CPD, buying my books, taking my online courses, listening to my podcasts, reading my newseltters, or supporting what we do at Eedi. This feels like a good way for an old man to say thanks.

The future of the Ultimate Retrieval Tool

I reckon there will eventually be around 800 skills. So far, I have built 260. I started in the Number section, and I am working my way down, adding new skills every week — so keep checking back.

I also have plans to improve the interleaved function and add a smart spacing element. That’s my treat for the next few weeks

I am not putting this on the home page of my website just yet. I wanted to give it to you, my dear newsletter readers, first — a thank you for putting up with all my geeky AI posts over the last few months. Have a play around with it. Let me know in the comments what you like, how you have used it, and any suggestions for improvement.

The link to the tool is here.

One small ask…

You made it to the end, so I will keep the ask short. If you find this tool useful and can spare 30 seconds, please fill out the survey I mentioned at the top of this email — it genuinely shapes what I build next.

Thanks so much for reading. I really hope you and your students find this tool useful, and I hope you have a great week.

Craig

Every other guest in this series has been a researcher, a data scientist, or someone building AI tools for education. Gráinne Hallahan is none of those things. She is a former English teacher from Essex who now runs the questions and data at Teacher Tapp – the app that asks over 10,000 teachers in England three questions a day about what is going on in their schools.

And that is exactly why I wanted her on the show. Because we can debate the promise and peril of AI in education until we are blue in the face, but at some point, we have to ask: What are teachers actually doing? What are students actually doing? And does any of this match what the researchers and the builders think is happening?

It was an absolute pleasure to have Gráinne 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. 58% of secondary teachers have caught a student cheating with AI in the past month – and the real number is almost certainly higher

Teacher Tapp asked secondary teachers whether they had suspected a student of using AI in the past month, even though they had been told not to. 58% said yes.

That number is striking… but I think it is a significant underestimate.

I was in a school recently where a teacher told her students not to use AI for homework. I was chatting to some Year 10 lads, and they were quite proud of their system. They would screenshot their online homework, send it to ChatGPT, and get the answers. But they knew that if they all suddenly started getting 10 out of 10, the teacher would be suspicious. So they planned it out in advance: one would get question 3 wrong, another would get question 7 wrong. Different mistakes, different scores, no detectable pattern.

That is not an outlier. Kids are onto this. The question is what we do about it.

Of the teachers who did catch students cheating, 77% just spoke to the student, 35% asked them to redo the work, and only 13% issued a sanction. The sanction, assuming it’s sufficiently unappealing to the student, is likely to provide the most significant incentive… but perhaps not the one we hope for. It is likely to incentivise the kids to put more effort into hiding the fact that AI is doing the homework for them, just like the Year 10 lads did.

I believe there’s only one way to solve this problem, and it’s the argument I made in my Is Maths Homework Dead? piece and in my conversation with Adam Boxer. We have to stop judging students on the output they produce outside of class. Instead, set the homework, encourage them to use whatever support they want, and then test them in class on paper with no technology. Use the same questions, but ask for full working out. If they ace the test, then they learned from their homework, whether they used AI or not. If they bomb it, they did not, and can be held accountable for that. The purpose of homework has flipped from assessment to preparation for assessment, so the incentive to cheat disappears immediately.

Gráinne agreed but added that it also needs to be tied to a broader message: the reward is the work itself. She cited recent essays by university graduates who had used AI throughout their degrees and now felt they had cheated themselves out of the education they had paid for. You need the strategy and the culture shift together.

2. Four in ten teachers have never taught students about AI and do not plan to – and they might be right

27% of secondary teachers and 12% of primary teachers have taught students how to use AI tools this year. But the biggest group – four in ten – said they have never done it and do not plan to.

Is this a problem? Maybe, but maybe not.

The best AI tools are paid. The free ones are not great at things where precision matters (eg maths). I wrote about this recently: the workflows I demonstrate in my newsletters rely on Claude’s Max plan, which costs me £75 a month. When teachers try to replicate them on free ChatGPT, they get a completely different – and often terrible – output. And if that is true for teachers, it is even more true for students. Pointing a child towards a free chatbot that cannot count, hallucinates freely, and has access to all kinds of content you would not want a 12-year-old to see is not obviously the right move.

Gráinne drew a comparison I liked: we do not tell students to go home and use VR headsets because not all students have them. AI is not like that on the surface – in theory, everyone has access – but the quality spectrum is enormous, and that gap maps onto existing inequalities.

There is also the question of whether novice learners can even use AI effectively. I keep coming back to this. The only reason I can use AI productively is because I have enough subject knowledge to evaluate the output, catch the errors, and push back when it is wrong. A Year 7 student who does not know whether 4.2 × 19 is closer to 80 or 800 – to borrow Barbara Oakley’s example from earlier in this series – is in no position to catch a chatbot’s mistake.

So maybe those four in ten teachers are making a reasonable call. Their time is better spent teaching their subject than teaching a tool that changes every month, and that their students may not be able to use well anyway.

3. More teachers are using AI to plan lessons – and the ones who need it most are the ones most at risk

The proportion of teachers using AI to plan lessons has risen from 14% to 20% this year. Teachers with the least experience are the most likely to be doing it, which makes sense as they do not have a bank of lesson plans to draw on.

But Gráinne shared a story that should ring alarm bells. A teacher had been given geography on their timetable and was using AI to generate quick starter questions. A colleague walked past, looked at the board, and quietly pointed out that the answers were completely wrong. The teacher had no idea, because they did not know the subject well enough to spot the errors.

This is exactly the scenario that Adam Boxer warned about in his episode. A non-specialist or a novice who generates a worksheet on a topic they do not understand deeply is in a worse position than one who downloads a static resource from a trusted source, because at least with the static resource, someone with subject expertise created it.

Dan Meyer made a similar point in his episode: the best software engineers will be multiplied 10x by AI, and the least experienced will be multiplied by 0.9. An experienced teacher using AI to adapt a curriculum for their specific class is powerful. A novice teacher using AI because they do not know what else to do is dangerous.

Gráinne made a point here that I think deserves more attention. She talked about the relationship between a new teacher and their mentor, and asked me about my own NQT year – about Geraldine, my second in department, who watched my lessons, gave me feedback, told me when I was improving, but also taught me about staffroom culture, what to do and not do, how to survive, how to be a better person.

Gráinne’s argument: for novice teachers, that mentoring relationship is irreplaceable. AI can be a useful supplement, but it should not be the dominant resource they turn to. If we have a problem with access to mentors, we fix the mentoring problem. We do not try to solve it with AI.

4. 17% of schools have no AI policy at all – and in schools that do, teachers often do not know it exists

In secondary schools, 51% of teachers said their school has no AI policy for staff. But when Teacher Tapp asked senior leaders the same question, the vast majority said: “Of course we have an AI policy!”

The gap tells a familiar story: a policy exists somewhere in a shared drive or an email chain, but it has not been communicated, discussed, or embedded in the way that, say, Keeping Children Safe in Education gets printed out and read by every member of staff every September.

Gráinne’s point is simple: you cannot hold a teacher accountable for something you have not told them. If a parent complains that their child’s report was generated by AI, and you have no policy that says teachers should not be doing that, you have no leg to stand on.

The challenge, of course, is that any AI policy specific enough to be useful will be out of date within months. Gráinne’s suggestion is to make it principle-based rather than platform-specific: never put identifiable student information into an AI tool, here are the approved platforms, here are the dos and don’ts, and so on. And then actually give staff time to read it, ask questions, and understand it – rather than burying it in an email.

As an aside, I suspect a fair few school AI policies were themselves generated by AI. Which rather proves the point.

5. Teachers who believe they are getting better are the ones who want to come back – and no AI can replicate that

Gráinne saved her favourite Teacher Tapp finding for last… and it has nothing to do with AI.

Every year, Teacher Tapp asks teachers towards the end of the summer holidays whether they are looking forward to going back to school. Gráinne wanted to know what the enthusiastic ones had in common. She filtered by age, experience, subject, phase, whether they had children at home – nothing explained the pattern.

Then she crossed it with a different question: Compared to this time last year, do you think you are a better teacher?

There it was. Teachers who believed they had improved: 47% were looking forward to returning. Teachers who felt the same or worse: 29%.

That was the only variable that produced the pattern.

Gráinne’s point is that this is the most powerful lever we have for teacher retention – and it costs nothing. A good line manager who says to you, “I can see you are calmer in the classroom this year, your workload seems more manageable, you handled that situation better than you would have last year” – that conversation changes whether a teacher wants to come back in September.

No AI tool can do this. No policy can mandate it. It is a human relationship, built on observation and trust, delivered in a corridor or over a cup of tea. And in a series where we have spent eight episodes talking about what technology can and cannot do for education, I thought it was the perfect note to end on.

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 are a teacher in England and you are not already on Teacher Tapp, download it from the app store. Gráinne reads every message that comes in and answers every question request personally.

This is the ninth 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

6. Dan Meyer: podcast and write-up

7. Bibi Groot: podcast and write-up

8. Becky Allen: podcast and write-up

I would love to know: does your school have an AI policy? Do you know what it says? And do you agree with those four in ten teachers who have decided not to teach students about AI – or do you think that is a mistake?

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?

Here are three blog posts I found interesting this week:

1. Choosing to Stay Human by Ethan Mollick

Podcast listeners will know that I am in the midst of a mini-series on AI in education, and one of the key tensions between my guests has been their views on whether giving students access to an AI in various forms will support or hinder their learning. Here, Ethan argues that AI undermines learning when it lets students surrender the hard cognitive work that learning requires, so the central task for educators is designing AI use that pushes learners to think rather than think for them.

2. Incremental Fact Fluency v2 by Dylan Kane

Warning: you’re getting two maths picks this week. Dylan argues that breaking math facts into incremental fact families, practised in isolation, then interleaved, supported by relational strategies and careful teaching, is a viable way to get reluctant students enough practice to reach fluency… and he shares his resources!

3. Don’t Teach Fractions. Not Yet by Dan Murphy

And a second one for my fellow maths geeks to finish. Dan argues that fractions make sense to students only when taught after a deep grounding in ratios, using precise language and a concrete-to-abstract progression that mirrors how mathematics—and children's understanding—actually develops

Have a great weekend!

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, including my new Topics page.

Here are three blog posts I found interesting this week:

1. Non-Examples by Jamie Clark

There are a few things in life I love more than a well-crafted non-example. Jamie takes a deep dive into their features and uses, arguing that such near-miss contrast cases planned in advance, sharing surface features with examples but missing the critical defining feature, are what draw clear concept boundaries and prevent the misapplication that examples alone cannot.

2. “Screen Time” Needs a New Tune by Dan Meyer

“Nuance” is an overused word to describe someone’s views… but I don’t know of a better one to describe Dan’s writing on Edtech. Here, he argues that the call to ban (or reduce) screens in schools is too blunt. Screens are most valuable when teachers weave them in briefly and purposefully alongside talk, paper, and peers rather than letting them dominate the lesson

3. Teachers Deserve Better Access to Educational Evidence by Jim Hewitt and Nidhi Sachdeva

I love research. I love AI. Therefore, I love this. A new AI-powered tool, The Evidence Checker, gives teachers direct and nuanced (there we go again!) access to educational research so they can evaluate classroom claims for themselves rather than relying on intermediaries.

Have a great weekend!

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, including my new Topics page.

Once upon a time

When ChatGPT first arrived, I couldn’t believe my eyes. You could name a topic — adding and subtracting fractions, say — and out came a worksheet that looked half-decent. I made a lot of those worksheets… and learned absolutely nothing in the process.

That is trap number one: cognitive outsourcing, to use Paul Kirscner’s phrase. The AI does the thinking, you copy and paste, and your own pedagogical reasoning slowly atrophies.

When I eventually realised what I was doing, I overcorrected.

My ego was the issue. Twenty years in the classroom and ten years of reading and writing about variation theory had convinced me that no AI was going to teach me anything about question sequencing. So I would design every worked example, every sequence, every variation myself, and use the AI only to format it nicely. Pretty layout, neat margins, well-formatted equations, done.

That is trap number two. It’s nowhere near as serious as the first because you are doing all the thinking. But it robs you of an opportunity to use AI’s potential to help you become a better planner and resource creator.

What I am doing now sits in between. I use AI as a thought partner.

But what the heck does that mean? Well, in this newsletter, I’ll try to explain, using the framework that I employ and concrete examples from my own resource creation over the last week.

My five-step framework

So how do you turn an AI from a solo worker or an order-taker into a thought partner? There are five steps I follow.

1. End every prompt with “Before you build, ask me some clarifying questions”

Tag this onto the end of any prompt, and the experience changes. You are signalling that you do not want passive execution. You want feedback.

So, the prompt becomes: “I want a worksheet on adding and subtracting fractions. Before you build, ask me some clarifying questions.”

The response will not be a worksheet. It will be a series of questions that clarify in the AI’s mind — and, more importantly, in your own mind — what exactly you want. The sub-skills covered, the difficulty, the progression, the sequencing, the style of questions.

This is the easiest move in the whole framework. You can do it on your next prompt without changing anything else.

2. Invite the AI to be a critical collaborator

Step 1 will get you good clarifying questions. Step 2 will get you a peer.

At the start of your next prompt, define the AI’s role and the rules of engagement:

“You are an expert on maths pedagogy. I am inviting you to a co-construction session to design a high-quality teaching resource for Year 8s to practice adding and subtracting fractions. I don’t want you to just accept every suggestion I make. I want you to challenge me with reasons when appropriate. Here is what I am thinking… [tell the AI what you are thinking]… Before you build, ask me some clarifying questions.”

The crucial line is the one in the middle: “I don’t want you to just accept every suggestion I make. I want you to challenge me with reasons when appropriate.”

Without that, the AI’s default is to agree with you. With it, you give the AI permission to behave like a peer — to push back when it sees a problem, to flag a misconception you are about to plant, to suggest an alternative where one exists. But crucially, not just to disagree for the sake of disagreeing (we’ve all worked with people who do that, right?).

3. Scrutinise the first output the way you would scrutinise a colleague’s

Once you have answered the AI’s clarifying questions and it has produced its first version, treat it like a draft a colleague has shared in the office. Pick out what works and why. Pick out what does not and why. Suggest specific alternatives.

Something like: “I liked your choice of denominator for question four — it reduces the arithmetic burden, so the student’s working memory can stay on the concept of adding fractions, not on the arithmetic. But your jump from question five to question six is too steep. A better question six would be [X]. What do you think?”

The “and why” matters — vague feedback gets vague rewrites. So, does the “What do you think?” — it once again reminds the AI of its role, not as a passive producer but as a collaborative peer.

4. Ask for a second version. Then a third. Then maybe a fourth.

The second version will be much better than the first. But it is rarely finished. Expect another round or two of feedback before the resource is where you want it. This is normal. This is also where the real cognitive work happens — for both of you.

When I do this, the final resource often ends up better than what I would have produced on my own. Partly because I am drawing on the AI’s expertise. But mostly because I am freed from the grunt work — the typing, the formatting, the layout — and that frees up working memory to notice things about the pedagogy I would have missed if I had been doing the whole job myself.

5. At the end, ask: “What did you learn from this collaboration?”

This is the one most people skip. After a long, back-and-forth, with a well-deserved, lovely resource at the end of it, the temptation is to close the chat and get a coffee. Don’t. Ask the AI what it learned that it could apply to future collaboration sessions. It will go back through the conversation, look at what you praised and pushed back on, and return with a set of principles.

Save those principles. Paste them into the start of the next chat. The next resource will be better, quicker.

To take this to the next level, apply the same trick to yourself. After a resource is finished, ask yourself what you learned. Update your own running document. The two documents, yours and the AI’s, should compound across resources.

Before the examples: one more thing on prompting

Before we get into the nitty-gritty, one more thing.

I no longer type my prompts and responses. Typing on a keyboard is so March 2026.

I now dictate my prompts and responses using a tool called Wispr Flow, which transcribes voice into any text field on my computer. I mentioned this on my recent podcast with Becky Allen, and as a result of people clicking the link, I’ve got more free months of Wispr Flow use than science would suggest I’ll be able to use in my own lifetime.

There are two reasons Wispr Flow has changed how I work.

One, it is quicker than typing. Much quicker.

Two — and this is the bigger one — it frees up attention. When I am typing, a chunk of my working memory is on the keyboard. When I am dictating, all of it is on the idea. I can chase a half-formed thought before it slips away. I can include the kind of throwaway context — “what worries me here is…”, “and I think the reason is…” — that I would strip out if I were writing. The more I read about how large language models work, the more I realise they are hungry for that kind of context. The flowery asides we filter out when typing turn out to be exactly what sharpens the output.

So whenever you read one of my quoted prompts below, picture me talking into my computer, while my wife once again sighs, knowing that I am chatting to my best mate, Claude.

Five examples from last week

Strap yourselves in, we are about to get super nerdy. You have been warned

This week, I’ve moved on from my Shedloads of practice resources (in case it is useful, my last few builds were surface area and volume, area and circumference of a circle, representing linear inequalities on a graph, and an epic ratio page, along with this circle theorems interactive tool) to experimenting with designing teaching and testing sequences for categorical atoms.

The idea comes from my good friend, Kris Boulton, over at the excellent Unstoppable Learning Substack, which itself builds on the pioneering work of Siegfried Engelmann on direct instruction and logically faultless communication.

I will spare you the full theory (Kris is the person to read on this, starting with this post). The short version: when you want to communicate a categorical concept precisely — what counts as an integer, what counts as an exterior angle, what counts as a valid set of angles on a straight line — a carefully chosen sequence of non-examples and positive examples lets students induce and encode the rule for themselves, without you ever stating it.

I have been writing these sequences by hand for three or four years. I find them intellectually stimulating. I also find them hard. So when AI’s diagram capability finally cracked (I wrote about this here and here), I wondered: could AI help me build the animations needed to draw students’ attention to the minimal changes between positive and negative examples, and help me sequence these positive and negative examples in the first place?

What follows are five exchanges from recent builds. The first three are from an “Is this an integer?” sequence. The fourth is from an “Are these angles on a straight line?” sequence, and the last from “Is this written in standard form?”. All five exchanges shaped the final resource in ways I would not have got to on my own… and that the AI would not have got to on its own.

Exchange 1: How many examples do we need? (the AI pushes back)

I had drafted a sequence of five examples to try to convey to students as clearly as I could: “What is an integer?” My prompt to Claude was:

Perhaps we start off with a decimal like 0.2 for our first non-examples, and then show that transitioning to an integer, 2.

For our second positive example, we show a really big integer, to try and push the boundaries of it, so something like 53,076, or something crazy like that.

For the third positive example, we need to show them a negative number, so maybe something like -3, and then we could transition that into -1/3 for the final non-example.

Would that cover everything, or do we need to also show them that a negative decimal is not an integer? I think the five examples might be enough on their own. Let me know what you think.

What I expected was: “Sounds good, Craig, you are a genius. Let me draft the sequence.” What I got was:

I’d push for adding -2.5 as Example 6.

After Example 5 the student has seen one positive decimal as N (0.2) and one negative fraction as N (-1/3). To classify -2.5 as not-an-integer they need to combine two threads from the sequence — ‘decimal means not whole’ and ‘negative is OK only when whole’ — and apply both to a single new case. That’s the kind of inference that works for the strong half of the class and fails for the rest.

A KS3 student who has internalised ‘fractions like -1/3 aren’t integers’ can still trip on -2.5 because decimals feel categorically different from fractions at that age, even though we know they’re the same thing.”

I had to stop and think.

The argument was right. I had been reasoning from my own perspective — for a teacher who knows that -2.5 is just another not-an-integer, of course, five examples is enough. But the sequence was not for me. It was for a Year 8 student in the bottom set, who may have noticed that decimals are not integers, and negative fractions are not integers, but negative whole numbers are. They have not yet been given any signal as to which category negative decimals belong to.

We went with six examples.

The bit that mattered was not the conclusion. It was the pedagogical reasoning that the AI shared. Claude did not just disagree — it surfaced something that my curse of knowledge blinded me to.

Don’t say I didn’t warn you: this post is for the super geeks.

Exchange 2: Two ideas I really wanted to use (the AI pushed back again)

Once we had agreed on six examples, I got excited about the closing transition.

My first idea was to use -1.3 as the sixth example, transitioning from -1/3 in the fifth. The reason: a really slick animation. With -1/3 written as a stacked fraction, you could fade out the dividing line, slide the denominator down, and morph it into -1.3. Minimal change. Visually clean. I was pleased with myself.

Claude was not.

I’d reject this on a stronger ground than ‘no mathematical relationship.’ A student who watches -1/3 morph visually into -1.3 — bar vanishes, denominator slides up to become a decimal digit — could plausibly encode ‘so -1/3 must equal -1.3’ as a takeaway. That’s an active misconception we’d be planting in service of a slick animation. Bad trade.

Claude doesn’t beat around the bush.

But Claude was correct. I had been so focused on the visual elegance that I had missed what the student watching it would actually take away.

So I tried again. My second idea was to choose a negative fraction and pair it with its decimal equivalent — same value, two surface forms, both not-integers. Pedagogically tidy. Again, pushback:

I’d skip this. The pedagogical payoff is ‘watch one number in two surface forms, both not-integer,’ but the student has to compute the equivalence to receive the payoff — recognise that -3/2 = -1.5. If they don’t, they see two different examples and the equivalence does no work. For an atom whose entire job is integer/not-integer, asking the student to do parallel fraction-to-decimal arithmetic distracts from the main question.

Two ideas, both rejected. Both, on reflection, rightly.

I had been thinking about what looked good and what was conceptually clever. The AI had been thinking about what the student would actually perceive and what cognitive load would be wasted. The student-side reasoning was the right reasoning all along — and I had missed it twice in a row.

Exchange 3: My turn to push back (the negative sign needs to physically stay put)

The first two exchanges could make the AI sound like the wise sage. So here is one going the other way.

We had settled on the closing block of the integer teaching sequence:

I needed students to notice that the negative sign had three different things after it, resulting in two different answers. The minus is the constant. What follows the negative sign is what does the work.

But the way Claude had set up the animations, the minus sign was redrawn every time the example changed and moved based on the length of the number that followed it — not held in place as a fixed anchor the student could rely on. This reduced the likelihood that students would notice it.

This might look like a design quibble. It was not.

Claude, I’m not happy with the transitions through the closing block. The whole pedagogical point of E4, E5, E6 is that the negative sign is the invariant — same minus, different operand. But if the minus sign on the screen disappears every time the example changes and gets redrawn, the student cannot see that. The negative sign has to stay in the exact same position throughout. Visibly. The whole time. The thing-after-it can morph or fade, but the minus does not move. And I think this is a general principle we need to bake into the handoff document: any constant feature that survives a minimal change must remain in its exact same position and visible at all times.

The reply was a clean concession, with a useful refinement:

You’re right and the argument is strong. The whole point of a minimal-change transition is that the student can read off what changed by reference to what stayed the same. If everything in the candidate visually leaves and gets replaced — even by an identical character — the student loses the anchor that lets the change land as minimal. One refinement: pieces should be identified by pedagogical role, not by character coincidence. The ‘3’ in ‘-3’ and the ‘3’ in ‘-1/3’ share the character but not the role — operand body in one, fraction denominator in the other. Linking them would require motion to morph one into the other, which isn’t invariance — it’s an additional animation. The minus is the right invariant. The bodies of the operands should cross-fade as wholes.

That refinement mattered. I had been about to ask whether the ‘3’ should be linked across examples too? But Claude was right. The ‘3’ in ‘-3’ and the ‘3’ in ‘-1/3’ are different mathematical objects, even if they share a glyph. Treating them as the same piece would have risked planting a quieter misconception.

What might look like a fiddly design decision was, in fact, the load-bearing pedagogical move of the whole closing block. The negative sign, held still on screen, tells the student: this is the rule. Without it, the three examples could read as unrelated, with reasons the student has to guess. With it, they read as one rule applied to three operands.

Would I have reached this conclusion on my own? Maybe. But perhaps my attention and working memory reserves would have been depleted with all the other aspects of the sequence I had to contend with. Whereas with Claude as a thought partner, some of those reserves were freed up, enabling me to spot things like this

Exchange 4: How much do we need to teach?

So far, three exchanges from the same number-focused build. As regular readers will know, I’m currently obsessed with making diagram-heavy resources using AI. So, here is one from a geometry resource: “Are these angles on a straight line?”

The critical feature for this sequence is that the marked angles must tile exactly half a plane on one side of the line. Two of the ways this can go wrong are:

1. Overfill: the student includes an extra angle that spills past 180°

2. Gap: the student marks angles that fall short of 180°, leaving a gap before the line

Both are failure modes of the same rule. Both produce incorrect answers for similar reasons. So the question was: do we include both in the example sequence — two closing non-examples, one of each — or do we teach one and trust students to generalise to the other?

Claude proposed the latter. Include one of them in the teaching sequence (the one students are less likely to spot on their own), and give both equal weight in the testing sequence that follows.

The reasoning was about scope. The teaching sequence is finite. Every example has a cost. Some failure modes need to be shown explicitly because students will not get to them on their own (like the negative decimal from the integer sequence above); others will land once the rule is understood, and adding an example for them just bloats the sequence without paying off. A critical feature does not require a teaching example for each failure mode. It needs a teaching example for the failure mode that students will not spontaneously generalise to.

I would not have arrived at this distinction on my own. I would have over-engineered — teach everything explicitly, leave nothing to chance. The AI’s framing — teach what they cannot generalise, test what they can — is sharper. And the resulting sequence is shorter and better.

The principle might still be wrong, of course. Maybe students do not generalise to overfill as easily as we think. Likewise, in the “Is this an integer sequence?”, maybe I could have got away with just five examples in the teaching sequence. The testing sequence will tell us. If gap performance is much worse than overfill performance after teaching, we add a gap example to the next iteration of the teaching sequence. That is the kind of judgment a thought partner helps you make. You then gather data from real students in real classrooms and adapt accordingly.

Exchange 5: When a non-example fails too hard

Anyone still with me?

Okay, one more example, this one is hot off the press, from a “Is this written in standard form?” sequence that I wrote last night.

Here is the teaching sequence Claude initially proposed:

Notice the final example. This is not in standard form for two reasons: the coefficient is out of range, and the exponent wasn’t an integer.

The intention was efficiency: get a final example in on both critical features at the close. The pedagogical effect, I realised, was the opposite.

We have a big problem with example six. We’ve already communicated to the students that standard form fails if the power is not an integer. Students may rightly look at example six and think, “I don’t even need to pay attention to that starting number because the power is not an integer, so it must not be standard form.”

Claude got it immediately:

The double-failure in E6 doesn’t just risk a “two negatives” rule-induction — it actively lets students off from processing the coefficient. They see the non-integer exponent and stop reading. The headline critical feature gets bypassed at exactly the slot where it should land hardest.

The deeper point that generalises is the difference between “this example fails on one thing” and “this example presents one new thing for the student to process.” These are not always the same. If you’ve already shown a student that a non-integer exponent disqualifies an expression, and you then show them another expression where the exponent isn’t an integer and the coefficient is also out of range, you’ve created a situation where the student can land on the right verdict without ever reading the coefficient. The critical feature you wanted to teach gets skipped, even though it’s there on the screen.

That’s the kind of thing I need to be thinking about when resource creating - and the very thing I might have missed if I had been bogged down with the minutia of creating the rest of the sequence.

At the end of each of these five exchanges, I asked Claude to reflect upon what he had learned and then write up the principles in a document that I could share at the start of every subsequent page build.

The resources

This is very much a work in progress. I have not added them to my Topics page yet. But in case you want to try out the resources with your students, here are the ones that I (well, me and Claude) have built so far, complete with animations:

1. Is this the perimeter?

2. Is this a square?

3. Is this a triangle?

4. Is this a factor?

5. Are these like terms?

6. Is this an exterior angle?

7. Is this an integer?

8. Are these angles on a straight line?

9. Is this written in standard form?

One last thing: Is this all a bit sad?

You might be reading this and thinking: Come on, Craig. This is just collaborating with a colleague… except the colleague is a computer, you weirdo.

My wife would agree.

And, look, I would not pretend that bouncing ideas off Claude is better than bouncing them off a great colleague with a mug of tea in the staffroom at the end of a long day. Of course it is not. A human colleague may have years of context on your students, your school, your department, and on you as a teacher. A human colleague can read the room. A human colleague gets the joke.

But here is the key question: How many of us, in our particular school, in our particular department, in our particular subject, have a colleague genuinely up for this kind of conversation — a back-and-forth on the pedagogical structure of a six-example sequence, with multiple back and forths, and who will do most of the grunt work? Oh, and who is available on a Sunday evening, or when the footy is on?

I am very lucky to know quite a few of those people. Most I met through writing and the podcast, including Kris Boulton. But even Kris is not available 24/7 (23/7, maybe).

Claude is.

That is not better than the staffroom conversation. But it is much better than not having the conversation at all.

Over to you

You made it to the end! You deserve a medal.

But instead, I have a challenge: If you have given AI a go as a thought partner — actually asked it “Before you build, ask me some clarifying questions”, actually asked it “What did you learn from this collaboration?” — drop your examples in the comments. The dynamic I have described above is the thing I most want to hear about. Not the prompts that worked the first time, but the pushbacks that made you stop, think, and learn something.

My wife thinks I have lost it. She might be right. But I have never had this much fun making maths resources.

Thanks so much for reading, and have a great week.

Craig

PS You can push AI too far. Here’s how Claude responded to one of my recent suggestions after an hour of back-and-forths:

AIs are getting increasingly human by the day.

Becky Allen trained as an economics teacher, taught some maths to fill up her timetable (she still feels guilty about how badly she taught it), and has spent the last 20 years as an education researcher. She co-founded Teacher Tapp, the app that asks over 10,000 teachers a day what is going on in their schools. She co-authored The Teacher Gap, which I think is one of the most underrated education books of the last decade. And she now spends most of her time at Alpha School, evaluating learning apps, writing specifications grounded in the science of learning, and working on assessment.

She is also one of the most genuinely optimistic people about AI in education that I have spoken to in this series. But her optimism has edges to it that I was not expecting.

It was an absolute pleasure to have Becky 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. Learning apps will replicate best practice at scale – but they will not discover new ways to teach

Here is where Becky’s optimism starts, and it is worth understanding precisely what she is and is not claiming.

Becky is claiming that AI will allow us to take what we already understand about effective teaching and build it into apps that can deliver that practice to every student, consistently, at scale. In a profession she describes as a leaky bucket of expertise – where we lose enormous numbers of teachers every year, struggle to explain to new teachers how they should teach, and watch them gradually learn through trial and error before many of them leave again – that is not a small thing.

Becky is not claiming that AI will figure out how to teach better than we currently know how. Why? Because education has a verification problem.

In chess, a machine can try a million moves and instantly verify which ones lead to winning. In driving, a car can try a manoeuvre and immediately check whether it crashed. This is what makes reinforcement learning – the technology behind most of AI’s breakthroughs – so powerful. Try something, check if it worked, do more of what works.

But in education, we cannot cut open a student’s brain and check what happened. We can run an immediate assessment, but it cannot measure everything. Crucially, a student can perform well in the moment and forget everything a week later. The things we really care about – deep understanding, flexible knowledge, transfer – play out over weeks, months, years. You cannot run reinforcement learning cycles on that timescale.

This is why, as Becky pointed out, the teaching profession itself has not cracked the problem of “Why do we not know how to teach everything by now?” We do not, because we have always had a verification problem. And so do the machines.

I found this a useful addition to Carl Hendrick’s prediction that AI-powered platforms would discover things about learning that we have been trying to figure out for decades. Becky’s position is more cautious: the platforms will be excellent at implementing what we already know, but the hard problem of verifying whether something actually worked for a particular student over time is not going away.

2. The best things AI can do right now are the things a great one-to-one tutor would do

I asked Becky what she has actually seen that impressed her – not promises, not prototypes, but things working right now with real students. She gave me two examples.

The first is prerequisite knowledge activation. Before a lesson on circle geometry, the app asks the student: “We are about to learn something new about circles, and in order to do this, we need to remember something called Pythagoras’ theorem. Just tell me anything you can remember about it”. The student responds in free text. If they remember everything, the conversation pushes further to find the limits of their knowledge. If they say they cannot remember anything, a follow-up prompt narrows it down: Pythagoras’ theorem is about a particular type of triangle, a right-angled triangle – does that ring a bell?

This is what a good one-to-one tutor would do. It takes about three minutes. And it is so much more productive than either a passive reminder or a single multiple-choice question.

I pushed back. In my experience, the best teachers do check prerequisite knowledge – either through a whole-class starter on mini-whiteboards or by analysing homework data before the lesson. Becky agreed that those are good approaches, but pointed out that the app version is adaptive in a way that whole-class checks cannot be. The conversation with each student can last three minutes or thirty seconds, depending on what they already know. In a class of 30, you cannot do that.

The second example Becky cited is guided worked examples. Instead of showing a student a complete worked example on screen – which Becky says students tend to scan with their eyes darting around rather than working through step by step – the app breaks it up. It shows the first part. Then it asks: right, so what are we doing now? The student says: dividing both sides by six. The app says: off you go, tell me when you have the answer. Then we move to the next stage.

I see the static version of this going wrong in classrooms all the time. A teacher puts a worked example on the board, maybe talks through it, and then the students copy it down. But as I wrote in my Myth of Copying Things Down piece, nothing magical happens in the process of transferring something from a board to a book. The app version forces the student to engage with each step before they can see the next one. That is a genuine improvement.

3. Khanmigo failed because students will always take the path of least resistance – and you have to design for that

Sal Khan stood on a stage and told the world that Khanmigo would revolutionise learning. As Dan Meyer pointed out in our conversation, Khan has since admitted it did not do what they hoped.

Becky was not surprised. She has a 15-year-old daughter who finds maths difficult. Becky spends a lot of time watching her do homework on Sparx Maths. Her daughter’s go-to when stuck is to rewatch the video. She could open ChatGPT and ask for help. She does not – not because she thinks it would be wrong, but because she wants the shortest possible route to getting the questions done. And rewatching a two-minute video is currently shorter than typing a question into a chatbot.

This is the behavioural science at work, and it connects to everything Bibi Groot told me about the research on unconstrained AI tutors. The Turkish high school study. The chess study. Students who had access to AI help did worse on the final test because they had offloaded the cognitive work. Students who could request hints whenever they wanted could not stop themselves from doing so, even when they knew it was hurting their learning. Michael Pershan’s line from the Dan Meyer episode keeps coming back to me: any tutor that the tutee can control is not a tutor.

Becky’s conclusion is the same as the one we arrived at with Bibi and Adam Boxer, but she puts it in a different way: learning platforms have to be highly controlled. Students should have relatively little autonomy in how they navigate the app, because when you ask children what they want to do, they do not choose the most efficient and productive activities for learning. At Alpha, the model is highly constrained. The AI decides what the student works on, in what order, and for how long.

This, Becky said, was why Khanmigo was largely a dead end. It gave students a chatbot. Students had to choose to use it. Most did not, because there was always an easier alternative. The system has to control the learning pathway. There is no way around that.

4. Alpha School is a benign laboratory – and that is exactly what it should be

I asked Becky the question that I know is on many people’s minds when they hear about Alpha School: can you learn anything useful from a school that charges top-of-market tuition and enrols children of wealthy, highly motivated parents?

Her answer was more interesting than I expected. She agreed the students would probably be successful regardless of where they went to school. But she argued that is exactly why you should be building there. When you are working in an uncertain environment with rapidly changing technology, you should build in benign circumstances to figure out what works without worrying that you are damaging vulnerable children.

The critics are right that Alpha raises unanswered questions about what happens when you try this with different types of children, families, and less well-resourced schools. But Becky’s point is that you have to answer the “can this work at all?” question before you answer the “can this work for everyone?” question.

I asked if she would send her own kids. She said no – but not for the reasons I expected. Her children are 12 and 15 and thriving at a large comprehensive. Her daughter has found her gang, sings in two pop bands. You cannot do that in a school of 50 students. The smallness of Alpha is, for Becky, the weird part – not the screens, not the lack of traditional teachers, but having so few peers to find your niche among.

I asked what Alpha has got wrong. She said there is a gap between the two-hour morning learning block and the afternoon project-based activities. For students who love a subject, there is no space to go deeper – to learn about the history of maths, or how a scientific discovery was made, or to explore ideas beyond the core curriculum. She thinks they will figure it out, but it has not been figured out yet.

5. Mainstream schools will mostly stay the same – but marking will get solved and micro schools will appear at the margins

I asked Becky to make a prediction. Her answer surprised me, because after everything she had said about how optimistic she is about learning apps, I expected a bolder vision for how schools would change.

She does not think mainstream schools will change much. She wrote a whole book about this during the pandemic – The Next Big Thing in School Improvement with Matt Evans and Ben White – and her argument is that schools are complex systems that cannot simply adopt a new operational model overnight. You do not have the infrastructure, the buildings, the adults, or the political will to ship Alpha’s model into a local comprehensive.

What she does think will happen is that learning platforms for independent study will get much better. Marking will get solved – not eliminated, but reduced to a point where a teacher looks at a few scripts, the machine handles the rest, and the teacher gets a summary of whole-class feedback to deliver. Given that Teacher Tapp data shows overwhelmingly that teachers dislike marking, this alone would be a meaningful improvement in quality of life.

And at the margins, she sees micro-schools appearing. Small schools of 50 to 100 students, funded initially in the private sector and as alternative provision, offering a viable option for the 20% of students classified as persistently absent and the one and a half percent in elective home education. These are children for whom mainstream secondary school is not working – not because they are incapable, but because they cannot deal with the environment. Software-powered micro-schools could give them a structured education in a setting they can manage.

Becky also made a point I had not considered. At the moment, if a school shrinks below a certain size, it can no longer afford subject specialists in every area. But with learning apps, a student could be the only person in their school taking psychology or Further Maths and still receive a high-quality course. You end up with a mixed economy: the traditional school model for most students, supplemented by online provision for niche subjects where the numbers do not add up.

Adam Boxer predicted that AI would not revolutionise education because nothing ever does. Becky agrees – at least for existing mainstream schools. But she sees something different growing at the edges. Not a revolution. More like a slow, quiet expansion of what is possible for the students the current system cannot serve.

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 Becky’s Substack (Prof Becky Allen), her assessment-focused Substack with Matt Evans (100% Assessment), and of course Teacher Tapp if you are a teacher in England and not already on it.

This is the eighth 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

6. Dan Meyer: podcast and write-up

7. Bibi Groot: podcast and write-up

I would love to know: Does Becky’s vision of micro-schools appearing at the margins resonate with you? And do you agree that mainstream schools will mostly stay the same, or do you think the pressure will build faster than she expects?

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?

Here are three blog posts I found interesting this week:

1. How Much Practice Do Students Need? by Dylan Kane

As much as I love research into retrieval practise, nothing beats reading about how an actual teacher is using it day to day with their students. Dylan argues that procedural skills in maths should be practised in short, spaced chunks across multiple days — guided by daily checks for understanding — until students can reliably retrieve them, because durable learning depends on building storage strength, not just performance within a single lesson.

2. The Origins of Belonging by Jon Hutchinson

Powerful stuff from Jon here. He argues that belonging is best understood not as a soft moral goal but as a cognitive one — pupils who are uncertain whether they belong spend working memory on social vigilance that should be available for learning, so building belonging directly expands the cognitive bandwidth available for the content we teach.

3. Predicting the unpredictable: Can mathematics teachers learn to anticipate student errors? by Colin Foster and Aidan Stannard

Okay, a bit of a self-indulgent selection to end this week, as it involves my Diagnostic Questions website, but I hope you’ll forgive me because it’s fascinating. Colin and Aidan show that secondary maths teachers can rapidly sharpen their ability to predict which errors students will make — and in what proportions — through short, repeated cycles of prediction against diagnostic question data, making this a promising, low-cost route to developing a critical strand of Pedagogical Content Knowledge.

Have a great weekend!

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, including my new Topics page.

If you want to know why I think this matters, read on below!

Another week, another load of diagram-heavy maths resources created. I started with the full suite of statistical diagrams I was missing, dabbled in some geometry topics, treated myself to a bit of graph plotting, and then I had a proper go at the big one: circle theorems, which felt like the holy grail because of the complexity of the diagrams.

Here is a flavour of the fruits of my labour.

A full suite of carrying out and describing transformations:

Drawing and Interpreting Box Plots:

Plotting straight lines and quadratic graphs:

And, in what I am calling the greatest achievement of my life to date, a full suite of Circle Theorem resources.

I’ve built each resource to perfectly render as a worksheet as well:

Now, you could look at all that and reasonably ask: so what?

Maths, of all subjects, is the one with the most high-quality, freely available resources. Textbooks. PowerPoints. Worksheets. Whole websites that have been serving teachers for decades. Why do we need another bundle?

Two reasons. And I think both matter.

1. These resources are not static

Take a worksheet with 10 questions on angles in a triangle. Your class smashes through them. Five students need more practice. What do you do? Hunt for another worksheet? Make one up on the board? Move on and hope?

With AI-generated content, you can just generate more. 5 more, 20 more, 40 more. Same topic, same difficulty, or scaled up. Complete with answers and working out, Click.

But the quantity is not the interesting bit. The interesting bit is the ability to combine.

Here is what I mean. This is a screenshot of the top of my circle theorems page:

You can pick the exact theorems you want students to practise. Imagine that you have just taught the angle at the centre and angles in the same segment. Students have practised each skill in isolation, but now they need to get good at switching between the two. Tick those two boxes. Generate as many questions as you want. Then, once you have taught the next theorem, add it in. Then the next one.

This allows you to provide students with focused practise on any individual skill, as well as interleaved practise, ensuring they have the opportunity to develop the all-important skill of discrimination, which research suggests is crucial.

That is what is new. Not the volume of questions. The fact that the mix is in the teacher’s hands.

2. It opens the door to resource types I could not build before

My Shedloads of Practice resources are designed to do one thing: consolidate. Drill the skill, secure the procedure, practise switching, get the reps in. That is useful. But it is not the only thing students need.

So, I have started experimenting with a new resource type called Variation 6. It leans on the work I have been doing for the last seven years on variationtheory.com.

The structure is constrained: six questions only. Each one draws students’ attention to something specific. What changes between questions? What stays the same? How far can we push the concept? What happens if we reverse?

Here is one I built earlier for my 7-year old, on adding a single-digit number crossing the 10 boundary:

But because AI's ability to create high-quality diagrams has increased, this opens the door to doing more complex diagram-heavy topics such as this

My next project is to build a lot more Variation 6 resources, including for topics that were locked off when AI could not draw well. I am yet to break this news to my wife.

Building Variation 6 resources is not easy. It is not enough to write the prompt, click a button, go and make a brew, and come back to a finished resource. Variation 6 has been a proper back-and-forth with Claude. I sit there, sketch the structure, push back when a question or sequence isn't doing what I want, ask for tweaks, swap things in and out. I am using Claude as a thought partner, not a printing press. And I’m having a great time doing it.

I will write more about that process in a future post, sharing some of the back-and-forths I’m having, including examples where Claude does something daft, and when I do something daft and how Claude massively improves it, because I think it matters. The output is only as good as the conversation you have to get there.

In the meantime, the best place to find all the latest resources I’m working on, mapped to individual topics, is on my Topics Index page.

Pulling it together

So why does this development in AI’s ability to draw matter? Not because we are short of maths resources. We are not.

It matters because we can now produce diagrams in the quantities and combinations that an individual class needs at any given moment. And it matters because it unlocks a category of resource — variation-led, attention-shaping — that we just could not build with AI before.

Both of these will land or fall on what teachers actually do with them. So I am genuinely keen to hear: if you have had a go at any of this, what have you been making? What is working? What is not?

Let me know in the comments.

Thanks so much for reading.

Craig

Bibi Groot started her career at the Behavioural Insights Team – the government unit that became famous for getting people to pay their taxes on time by changing a single line in a letter. She ran randomised control trials in further education colleges, earned a PhD in Behavioural Public Policy at UCL, had twins in 2022, and then joined Eedi as our first behavioural scientist because, as she put it, education is where her heart is.

I have wanted to get Bibi on the podcast for a while, partly because she is one of the smartest people I work with, but mainly because listeners have been asking me to go deeper into the Google DeepMind study that we published earlier this year. The one that made the news. The one that Daisy Christodoulou, Carl Hendrick, and others all held up as the gold standard of AI tutoring research. Most importantly, the one that was the subject of our first-ever (no doubt AI-generated) hype Tweet.

Finally, we have made it!

So here are five takeaways from our conversation that I think every classroom teacher and school leader will find interesting.

1. It took two years to prove that Eedi works – and the results only showed up after 12 months

Before we get to the AI stuff, I want to start with something that I think matters more than most people realise.

For years, I just assumed Eedi was effective. I had built it. I believed in the pedagogy behind it. Teachers told me they loved it. But we had no rigorous evidence that it actually improved student learning outcomes.

So we ran a randomised control trial. Twenty schools. 3,448 Year 7 students. Two full academic years. The study was independently evaluated by a team led by Professor Steve Higgins, who has been instrumental in developing the Education Endowment Foundation toolkit. We did not run the analysis ourselves. We pre-registered the study design and outcome measures so we could not cherry-pick the results. And we committed to publishing whatever we found.

Here is what we found. At six months, there was no clear difference between the Eedi group and the control group. At twelve months, an effect size of 0.17 – roughly two to three months of additional progress. At eighteen months, 0.46 – which is large for a platform like ours. And at twenty-four months, it stabilised at around 0.3.

Two things struck me about this. First, the control schools were not doing nothing. They were using Sparx, MathsWatch, whatever they wanted. So the comparison is not Eedi versus a vacuum. It is Eedi versus the normal crowded landscape of maths ed-tech.

Second – and this is crucial – the positive results did not show up for a year. If we had run a six-month study, we would have concluded that Eedi had no measurable effect on learning. Bibi made the point that a lot of ed-tech companies run short studies, find positive results, and announce them. But the file drawer problem is real: studies that find nothing are far less likely to be published. You end up with 10 positive papers and a hidden drawer full of 100 null results, and no one knows the full picture.

We also ran a second trial that found a smaller, non-significant effect. When we compared the two, the difference came down to implementation. In the trial that worked, teachers were using Eedi analytics to inform their classroom teaching – bringing the worst-answered homework question into the next day’s starter, for example. In the trial that did not work, only 33% of teachers were using Eedi in the classroom at all. Most were just setting and forgetting homework.

This aligns with something Adam Boxer has been saying for years: homework is only effective when integrated with classwork.

2. A constrained AI tutor outperformed expert human tutors on short-term transfer – and we think we know why

Now to the Google DeepMind study. This is the one that generated all the attention, and I want to be upfront about what it is and what it is not.

It is a randomised control trial with 165 students, run over seven weeks, comparing three conditions: Eedi’s existing static hints (you get a question wrong, you read a hint, you retry), a human tutor who chats with the student via text, and an AI tutor powered by Google’s LearnLM with a human in the loop approving every message before it reaches the student.

It is not definitive proof that AI tutoring works. It is a small first study. Bibi was clear about that.

Here are the headline results. On the retry of the same question, both the human tutor and the AI tutor achieved over 90% success, compared to 65% for the static hint. No surprise there – an interactive dialogue is harder to ignore than a sentence you can click past in two seconds.

But on the next question in the sequence – a different question testing a related but distinct concept, which is what we call short-term transfer – the AI tutor condition scored 66%, the human tutor 60%, and the static hint 56%. The AI tutor outperformed the expert human tutors.

That was not what we expected. We had hoped the AI would perform almost as well as the humans. We did not expect it to do better.

Bibi’s working hypothesis is this: the AI was prompted to be Socratic. It asked students to reflect on their own thinking, to identify where they went wrong, to work through the reasoning steps themselves. The human tutors, meanwhile – experienced maths teachers with years of practice – tended to diagnose the misconception quickly and fix it as fast as possible. They skipped the metacognitive step. One tutor said afterwards: “The AI is giving the kind of feedback I would like to give if I had more time.”

I think that line is the most important sentence in the entire study.

On safety: out of over 3,000 conversations, there were zero safeguarding issues. Five messages contained any form of incorrect maths – and those were mostly the AI misreading an image (saying a student selected answer A when they had selected answer B). As I pointed out to Bibi, if I gave a thousand answers and five contained a small mistake, I would be more than happy with that ratio. My own error rate as a classroom teacher was considerably higher.

3. Unconstrained chatbots make students worse at maths, not better

I pushed back on Bibi. How do we know we are not wasting our time building a constrained AI tutor when kids could just screenshot a question and send it to ChatGPT?

Her answer drew on two studies that I think every teacher should know about.

The first, from a Turkish high school, randomised about a thousand maths students into two groups. One group could use GPT-4 while practising for their exam. The other could not. Then both groups sat a closed-book exam with no AI access. The students who had practised with AI did worse. They had offloaded the cognitive work to the chatbot, and when it was taken away, they had less in their heads than the students who had struggled on their own.

The second study, by the same researchers (Bastani and Poultis), looked at chess practice. Students who could request hints whenever they wanted did worse than students who received tips at regular intervals but could not request them. The researchers found that students knew requesting help was hurting their learning – but they could not stop themselves, because struggling is unpleasant and humans are built to reduce friction.

This is behavioural science at work. Make it easy to get the answer, and students will get the answer. But getting the answer is not the same as learning.

Michael Pershan says that any tutor that the tutee can control is not a tutor. Adam Boxer went further in our conversation, saying schools should be communicating to students that using AI is a waste of their time. I am not sure I would go that far. But the evidence that unconstrained chatbots hurt learning is now strong enough that I think any school allowing students to use ChatGPT for maths homework should be worried.

The AI tutor we designed with Google DeepMind is constrained in five specific ways. It only activates after an incorrect diagnostic question. Students cannot send it screenshots of their homework. It addresses only the specific misconception revealed by the question. It uses guided questioning rather than giving answers. And it ends the interaction when the misconception is resolved, or progress has stalled. These are not arbitrary restrictions. They are designed to keep the cognitive work with the student.

4. The human tutors said the AI was like observing another teacher in the classroom

One of the things I was most curious about was how our human tutors – experienced maths teachers, remember – felt about having an AI suggest messages that they then had to approve or edit.

Bibi surveyed them at the start and end of the trial and found four archetypes. Enthusiasts who were all in from the beginning and stayed there. Pragmatists who saw specific use cases and stuck to them. Ambivalents who were unsure what to expect and still were unsure by the end. And sceptics – a small group who worried that their role was changing and were concerned about their future.

The sceptics are worth taking seriously. Carl Hendrick made the point in his episode that teachers seem to accept every other industry will be affected by AI but somehow think education is immune. That is not realistic. Things are going to change. The question is how.

But what struck me most was the feedback from the tutors who engaged deeply with the process. Several said it felt like observing another teacher in a classroom. Others said the AI came up with analogies or metaphors they had not thought of, and they took those ideas into their own tutoring practice. The AI was not replacing their expertise. It was extending it.

About 75% of the AI’s suggested messages were approved and sent without any edits. Of the 25% that were edited, roughly half were small tweaks – swapping “nearly” for “not quite”, removing emojis (the AI was very fond of emojis back then), trimming a three-sentence message down to two. The larger edits tended to come at the end of conversations, where the human tutor intervened to say: the student gets it now, let them go back to the lesson. The AI, prompted to be Socratic, kept asking questions. It could not read the social signals that the student was ready to move on.

Dan Meyer flagged this exact issue in his episode – that knowing when to stop intervening is one of the most sophisticated decisions a human tutor makes. The AI is not there yet. But Bibi described where Eedi is heading: from a human-in-the-loop (every message is checked) to a human-on-the-loop (the AI runs independently, with a human moderator flagged when conversations go off track). The human tutor is elevated from babysitter to expert intervener.

5. The next study will tell us what matters more: knowing the pedagogy or knowing the student

The Google DeepMind 2 study is running right now, as I write this, in 10 UK schools with 1,200 students. Results are expected by September 2026. And the question it is trying to answer fascinates me.

There are four arms. First, static content – the existing Eedi hints and videos. Second, a human tutor with no AI support. Third, an AI tutor armed with a strong pedagogy prompt – it knows about Newman’s error analysis, common misconceptions, effective questioning techniques – but knows nothing specific about the individual student. Fourth, an AI tutor with the same pedagogy prompt plus deep context on the student – their quiz history, the misconceptions they have struggled with in the past, how they have responded to previous tutoring interactions, what topics they are expected to struggle with next.

In other words: does it matter more to know how to teach, or to know who you are teaching?

Bibi put her prediction on the record. She thinks static content will come last. She thinks the human tutor and the pedagogy-only AI will be roughly on par. And she hopes the context-rich AI will come out on top – because if it does, it means we can build AI tutors that are meaningfully different from blank-slate chatbots.

I agree with her ranking, though I am slightly more bullish on the pedagogy-only AI beating the human tutor. How many teachers can diagnose misconceptions, think of several analogies, generate relevant follow-up questions, and find the time to work at the pace the student needs, in the hustle, bustle and demands of a busy classroom? That is essentially what our human tutors will be doing, supporting several students at once. The AI does not have these issues.

Bibi also flagged something from a Stanford study that I think is important: when an AI writing feedback tool was given access to student demographic data – gender and ethnicity – it gave Black students more praise and softer feedback than white students. The bias was baked into the training data. This is why Eedi’s AI tutor is given detailed information about what a student knows and does not know, but nothing about who they are. No name, no age, no gender, no ethnicity. Context on learning, not context on identity.

If static hints win, I am in trouble. But that is why we pre-register, publish regardless of results, and have an independent team running the analysis. It is also why this work matters. We are not building marketing material. We are trying to find out what actually helps students 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 Bibi’s posts on LinkedIn, where she writes about behavioural science, AI in education, and the work we are doing at Eedi around the world.

This is the seventh 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

6. Dan Meyer: podcast and write-up

We will be publishing the results of the Google DeepMind 2 study at the start of the next academic year. I am nervous and excited in roughly equal measure. If you want to follow along, subscribe to this newsletter and I will share them here first.

In the meantime, I would love to know: does the evidence so far change how you think about AI tutoring? And what question would you most want answered by the next study?

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?