The Science of Learning

A discussion with Carl Hendrick

Cognitive science tells us much about how our brains convert the blizzard of passing information into long-term knowledge and skills. Yet education systems often overlook it, while clinging to other practices that instruct less well. Why?

This question exasperates Carl Hendrick, a writer and professor of education who began his career teaching English at a state school in London. Today, he also advises Alpha School, the fashionable but controversial US-based private educator, hoping to teach kids in 2-hours a day, followed by social activities, with AI and no teachers.

To find out how (and if) this will work, we asked Carl what cognitive science tells us about learning, how Alpha operates, and why the decline of reading matters.

- Conor Griffin, AI Policy Perspectives

Conor: Many students hate memorization. You think it’s central to learning? Why?

Carl: Cognitive science tells us that a change in your long-term memory is the fundamental invariant of learning. It’s non-negotiable. Without it, you’ve learned nothing.

In the 1950s and 60s, we had a cognitive science revolution that helped us understand our memory as a system. In working memory, we have a hard constraint, whereby we can hold maybe five to seven elements at a given time. Our superpower is long-term memory which, as far as we know, offers an unlimited reserve to store things. If you store information in your long-term memory, then your working memory becomes outsized because you can effortlessly draw on the symbols and ideas you have stored. We’re doing it right now as we speak.

So cognitive science tells us that memorization is important to learning. But it also tells us that rote memorization of isolated facts is a debased form of learning, because memory is schematic—the concepts we learn need other concepts to stick to. So, if we want students to learn about World War Two, we don’t want them to just memorize dates. We want them to be able to build their own schema, or mental models, about the war. The question for educators becomes: how do you decompose World War 2 into atoms of knowledge that are teachable to a student, but which they can then rebuild?

Conor: Why are certain forms of teaching effective at this, like ‘retrieval practice’, where students are asked to recall facts or concepts from their memory?

In a word: friction.

The desirable difficulties framework, developed by Robert and Elizabeth Bjork, the psychologists and memory experts, says that learning is often counterintuitive. Things that feel like they are working to a student, such as re-reading and highlighting a text, often aren’t. And things that feel like they aren’t working, or are a struggle, like getting kids to dredge up something from their memory, often do lead to learning.

This comes back to how memory works. Human memory is not like a tape recorder. It is reconstructive. When we remember something, we seem to almost rewrite it. This is also why eyewitness testimony is so unreliable. But from a learning perspective, successfully retrieving a memory can strengthen it for the future.

With retrieval practice, the encoding of the knowledge seems to happen after the original event, when the students are trying to remember the thing. This is why the Bjorks see retrieval as a learning event, rather than a test of whether you have already learned something.

Conor: What’s the evidence that these cognitive-science interventions actually help students learn better?

There is a huge body of evidence from postgraduate students in labs, but little evidence for it in actual classrooms. And I say that as somebody who passionately believes in it.

A 2021 review found that we know almost nothing about how things like retrieval practice work with real kids and teachers. There is 70 years of evidence, but it is usually small studies with post-graduate students, focussed on maths and languages.

This is why we need to create new models of learning with AI. I want to see specific questions answered: “Given this third-grade reading problem in this state, what is the best intervention?“ It will be hard to isolate the effects in the way that we could with a vaccine. But with AI, we will be able to understand the effects of using a certain curriculum and sequencing the material in a certain way. We will be able to understand how to break down complex concepts into components to present to students progressively, without overloading their working memory.

Conor: Are there any interventions that do have solid evidence from real classrooms?

Yes, early reading and the Direct Instruction method.

For early reading, the debate is largely over. We know that instruction needs to be done with systematic, synthetic phonics. The brain is not built to read. So phonics teaches young children the individual sounds associated with specific letters, or letter combinations, before teaching them how to blend these sounds together to form words.

The most evidenced programme is Direct Instruction, a highly-structured method where teachers break complex concepts down into small, manageable steps. In the 1960s and ‘70s, it was subject to the biggest ever educational study in the US, Project Follow Through. This found that it was far more impactful for students than other approaches, for early reading, maths and other areas.

People tend to think of Direct Instruction as teachers just “telling kids stuff”. But instruction is actually the least important part of it. The most important part is the curriculum design. For example, there is a big focus on learning through contrasts. In the same way as we teach young children the colour red by showing them a red ball followed by a red triangle, you isolate the essential thing that you want them to know. This is a fundamental part of how we learn.

I think that there’s just no doubt that AI is going to be able to do this kind of curriculum design and sequencing far better than humans can. I think of the famous move 37 in AlphaGo. I think we’re going to see a similar thing for the design of curriculum and learning materials, where AI will just far surpass what humans can do.

How Alpha School works

Conor: That brings us to the new group of private schools in the US that you work with, that has big ambitions to use AI. You’re one of several prominent learning scientists advising Alpha. Why did you get involved?

Carl: Traditional schooling has been hugely successful; nothing has lifted more kids out of illiteracy and poverty. But there are problems.

The system is largely propped up by the 1-in-20 “superhero” teachers. The experience of pupils is massively heterogeneous. Since Covid-19, something else has changed. Many kids are not in school at all, or they are there so little that it’s almost a waste of time. Until now, educational technology, or EdTech, has also largely been a story of expensive failure because it creates digital versions of things that didn’t work in the first place.

But at Alpha School, we are designing apps where you get high-resolution data points from students at a phenomenal clip. You get correct/incorrect answers, but also things like latency and student hesitation. You can slice that data and make strong predictions about when and how to do interventions like retrieval practice. Or ‘spacing’—where students distribute their learning and practice over many lessons or days, rather than cramming it into a single session, to more productively engage their memory. This will allow us to see those interactive effects between the curricula and how you sequence the teaching materials.

Ultimately, I’m a materialist when it comes to learning. I think learning is governed by the same biological laws as digestion. If you can get the brain to pay attention to a certain sequence of information, and if knowledge is retrieved under a certain set of conditions, then I think learning will be almost guaranteed for 99% of kids.

Conor: At Alpha, how are kids supervised during this online learning?

Carl: The big change will be moving to AI coaching. This is where people will get uncomfortable because the answer is to mimic what a really good teacher does: warm/strict supervision. If you think about MOOCs—the massive open online courses that were much-hyped around 15 years ago—they were a failure because students weren’t accountable. Only 5 to 10 percent of people who were already highly motivated finished them.

Schooling fulfills the function of accountability. In the future, this will mean cameras on, and an AI monitoring student behavior, their latency, what websites they are looking at, their ability to focus and then producing a report for a human tutor. If you can get kids to concentrate on the apps, on the sequencing, phenomenal learning is going to happen.

But people won’t like this. You’ll see articles about “Orwellian spyware.” As a parent, I’d prefer this model to simply unleashing kids on the internet, crossing my fingers, and hoping for the best. AI will supervise better than humans. There are very skilled teachers who can supervise a class well, with eyes on the back of their head. But even then, there are kids who drift through the lesson, ‘cosplaying’ attention, and learning nothing.

Today, when you look at kids in a computer room, they might spend 20 percent of their time on a task. The rest of the time, when the teacher’s back is turned, they’ll be looking at websites. The model for AI should be a high level of accountability for students.

Conor: Students spend a couple of hours in the morning on this intensive online learning. What do they do in the afternoon?

They work on ‘life skills’. Broadly, longer-term projects, like setting up a YouTube channel, playing music in a band, or sport. Let’s not forget, Alpha is an expensive private school. So you will have kids there, who have had tremendous support from their parents, and so they will already be highly-motivated and able to do many of these things.

And so there is a question about whether that model is going to work elsewhere. But Alpha is trying to understand how learning happens. And they want to remove the standard model where kids sit at their desk for six to eight hours straight.

Conor: Alpha School doesn’t have conventional teachers. Even with a good AI system, won’t this mean a lack of social influence? With a good teacher, you want to impress them. How will you manage that?

Carl: We are in uncharted territory here, because we don’t have any data on this. In the Alpha model, the nearest thing they have is ‘Guides’ who receive a student’s data and have conversations with them, such as: “I noticed you went off-topic there for a while?”

That human question is a really important one and something that I’ve wrestled with. If more and more of the curriculum design, the instructional sequencing, and the assessment is delivered by AI, then it’s obviously going to come through tablets and screens. And then the question becomes: What do we lose?

There is still a social environment at Alpha. It’s a physical building. Kids come in. They have an assembly where they interact. And then, at nine o’clock, it’s two hours of online learning, although they’re still beside each other as they do this. There’s a system, where they have to generate a certain amount of experience points, and get a certain amount right, and so that brings some accountability.

But the question you’ve asked is a massive one, because it comes down to: What kind of schooling do we want for our kids? Are we going to eliminate something valuable—a teacher inspiring a kid? And I guess my answer is: we can’t scale that. We only hear the success stories, but what about the other 25 kids in that class? The Alpha model does offer scaling. You say to a kid, “Focus for two hours, hit these targets, and then you’re off to be a kid”. I think this will improve their academic outcomes. It might also improve their mental health. But these are difficult questions.

Conor: When one imagines the ideal schooling, it’s always the inspiring teacher. But the social influence of a teacher can also have an estranging effect, on those it doesn’t work for?

Carl: Exactly. One of my daughters was diagnosed with autism. School will be a challenge for her, because there is a noise in most classrooms, plus a lot of ambiguity and blurred edges. That can cause her distress. I was looking at an app that she’s using, where she flies through problems because there is more clarity and boundaries.

That was one reason I started working with Alpha. I realised that this will help kids with special educational needs that are not met in mainstream education. They could have a system where they’re making rapid progress, and not feeling like the bad kid in class all the time.

But again, the messaging is going to be very difficult. Critics will say you just want to “put autistic kids on screens”. But it’s for an hour or two a day. And then we want them to be kids.

Conor: What role do ‘AI tutors’ play in this vision?

Carl: There are three broad areas in education: curriculum (what we teach), instruction (how we teach it), and assessment (how we know they know). When the public hears ‘AI tutor’, I think they usually imagine a robot doing the instruction part.

But for me, AI tutoring is more about the back-end: curriculum design and assessment. If you accept that learning is an algorithmic enterprise governed by biological laws, then the most foundational elements are what you teach and how you assess it.

Designing a curriculum and teaching materials is more about engineering than about whether you can inspire kids. It’s about sequencing the materials. There are broader questions about the experience and social interaction that we want students to have. But, for more basic questions, like, “Is there a better way to teach quadratic equations?” I think we’ll see a shift from the current à la carte approach where teachers often decide what materials to use. And as uncomfortable as people are with it, I imagine that we will get AI-generated content that is just better than human-generated content.

Conor: You also see AI changing how we assess students?

Yes, your earlier point is right: if a kid writes something, they often want you (the teacher) to read it. But the Welsh educationalist Dylan Wiliam also referred to marking books as the most expensive public relations exercise in history. It’s tedious for teachers and much of it has zero impact on student outcomes.

This is also because human marking is unreliable. I was Head of Department of English and I saw situations where a student would do an essay, and get a D. And then it would be sent back to be re-marked and another person would give it an A. You can’t run a system like that.

The work on comparative judgment by Daisy Christodoulou and No More Marking, which calls for student writing to be assessed in comparison to other students’ work, rather than as an absolute, shows how AI could help improve the reliability of marking and reduce teacher workload.

But we also need to distinguish between the assessment of learning, and assessment for learning. Assessment shouldn’t just be an endpoint. It should be a regular temperature check. In a classroom, this kind of ‘checking for understanding’ is the most powerful thing a teacher can do, whether it’s a mini whiteboard session or a cold-calling activity. It’s imperfect, but it can identify misconceptions that students have.

The dream is that instead of teaching a pre-defined sequence of materials, an AI system takes regular temperature checks and operates like faders on a mixing desk. Based on the student’s response, the system identifies misconceptions in real-time and dynamically selects the best explanation for that specific kid.

This also means that we can meet the student at the point where the misconception occurs. Most current assessments are useless because the feedback loop is too long. If a kid writes an essay and gets it back a week later, they’ve already forgotten their thought process.

Conor: This would make education ‘adaptable’ to a given student. But you’ve also cautioned against using AI to match teaching materials to students’ different ‘learning styles’. What’s the nuance here?

Carl: The idea of learning styles emerged from a tradition in the 1960s and ‘70s, which thought about learning in an individualized way. As a teacher, I had to spend years trying to design lessons for different learner types, such as ‘visual learners’ or ‘kinesthetic’ learners, who purportedly learn better through hands-on experiences.

This idea has been empirically tested. Not only is there no evidence for learning styles, but it is the nearest thing we have in education to homeopathy or healing crystals. When you think about it, it’s preposterous to suggest that if you were teaching geography to an ‘auditory’ learner, you should make an audiobook to describe the continent of Europe. The content should determine the form.

Where there is evidence, is around the idea of dual-coding. Our brains process information through separate but interconnected channels: one for verbal information (spoken or written words) and one for visual information (images, animations etc.). There is value in designing explanations that think carefully about how to combine text and speech with visuals, so that you don’t overload students’ limited working memory.

For example, I see some great video explainers on YouTube, where, they show the actual differences between the size of the planets and it’s nothing at all like the maps of the solar system that you normally see. That’s an example of where the visual aspect really helps, but it has nothing to do with me being a visual learner.

Where education goes wrong

Conor: You are often supportive of explicit forms of instruction and skeptical of constructivist approaches where students explore and solve problems by themselves. Why?

Carl: Constructivism is a philosophy of meaning, not a pedagogy. It’s the belief that each individual constructs their own knowledge and understanding. It has its roots in Rousseau, but really got moving around the turn of the 20th century, with people like (the education reformer) John Dewey in the US, and (the Swiss psychologist) Jean Piaget in Europe.

As a philosophy, constructivism is true. We do construct meaning, but from an education perspective it has led to a belief in minimally guided instruction, the idea that you learn things best when you discover them for yourself, rather than having them taught to you.

What this does is to conflate the outcomes with the means. We do want students to be able to discover things for themselves. But discovery learning as a starting point for instruction is a disaster. Clear, sequenced instruction is effective for novices. And every kid is a novice in most domains.

If you’re learning to drive a car, you want explicit instruction up top—do this, then do that. Once a person is capable of driving, if you’re still doing that as a teacher, you’re getting in their way. That’s when the discovery element comes in.

There’s an ethical dimension to this too. Students who come from affluent backgrounds, with strong existing schemas of knowledge, who have been exposed to rich vocabulary around the dinner table, will flourish in an environment where there are few constraints or expectations in terms of their behaviour. Discovery learning privileges the already privileged.

Conor: This idea that novices can’t be expected to figure out things for themselves is also why you are skeptical of students learning by asking questions to chatbots?

Yes, the power of AI is in the monitoring, curriculum design, instructional design, and assessment. Novices won’t learn by discovering knowledge themselves by asking chatbots questions; they need explicit instruction.

Conor: Your argument that educators conflate the outcomes with the means, also explains why you’re skeptical of calls to teach students ‘21st century skills’, like critical thinking, creativity and adaptability, including as a response to AI?

Carl: Yes. Critical thinking skills and creativity are an outcome of systematic knowledge building, not a starting point that can be directly taught, as a generic skill.

Conor: The explicit instruction of knowledge that you advocate for might be effective for subjects like maths that can be more easily decomposed into sub-components. But what about subjects like English or history?

Carl: That is the Holy Grail of curriculum design. We have what AI pioneer Marvin Minsky and the computer scientist Walter Reitman would call “well-defined” and “ill-defined” domains.

Well-defined domains, like maths and science, are more clearly sequenced and hierarchical. They have concepts that are free-standing and can be taught in an isolated way. You can go back two steps if you miss something. They often have answers that are definitive or unambiguously correct. A lot of education apps that work well are in these areas.

In the humanities, knowledge doesn’t operate in that way. This will have implications for how we design instruction and assessment. But I go back to the materialist point. I believe these things are amenable to the laws of the universe and AI will discover what the optimum sequences are.

Why Reading Still Matters

Conor: You’ve spoken a lot about the decline of reading. Why does it worry you?

Carl: I find it sad that a lot of kids are not going to read Dostoevsky. But I also recognise that for me, growing up, reading was partly a feature of being bored, and we can’t pretend the conditions are the same today. It comes down to an ethical question: what kind of life do we want for our kids? This is not a question we can answer in a randomized controlled trial.

There are certain questions about reading that we can answer empirically in trials. One of the studies that stuck me dead in my tracks many years ago showed that the percentage of words we need to understand, in order to read a text, is 95%, which shocks everybody. The way that vocabulary is taught in schools is very poor, and inequitable, and there are almost certainly better ways to do it.

But we can’t just say to kids: “You need to read more!” That doesn’t wash. We need to answer the philosophical question of what we value reading for. The benefits are not always immediately obvious. They might pay off 20 years down the line. The literary critic Harold Bloom once said that Shakespeare shaped the modern consciousness, including our contemporary understanding of love, death, ambition. Something like the King James Bible affected English-language culture too. These were dominant texts, and culture passed through them.

What are we going to lose if we lose these books? I want my daughters to read because they’ll see that their struggles are not their own. So they can see other characters struggle too.