Soma for the Classroom

Why removing friction from learning removes the learning itself.

“Euphoric, narcotic, pleasantly hallucinant... All the advantages of Christianity and alcohol; none of their defects. Take a holiday from reality whenever you like, and come back without so much as a headache or a mythology.”

So Mustapha Mond introduces Soma in Aldous Huxley’s Brave New World, the state-sponsored drug designed to remove the discomfort from life. In the novel, people take it willingly and plentifully and why wouldn’t they? Who doesn’t want an easy life?

The desire to reduce friction is ancient and understandable and, indeed, throughout history, it has driven many great designs and innovations. Early bicycles had solid wooden wheels; the pneumatic tyre reduced the friction, both literally and figuratively.

In education, things are different. In that domain, the desire to remove friction is often counter-productive because it is in the discomfort that the learning happens. Unlike Odysseus’ Lotus-eaters, the people in Huxley’s world aren’t amnesiac or purposeless; they function perfectly. They’re just never made uncomfortable, and so are never changed by anything.

I suggested last week that we can’t simply trust engagement as a metric of educational value because it’s possible to engineer absorption in ways that de-couple it from learning.

This week I want to add a second mechanism to the first, because the two are easily confused and modern learning tools rely on both. This is the removal of friction, design that makes a task feel easy.

These mechanisms are not the same - one works by making you stay, the other by making you comfortable - and neither, on its own, has anything to do with whether learning has occurred. And it’s when they are combined that the trouble really begins.

Of course, the idea of frictionless learning, delivered by a machine, didn’t arrive with AI or even with the smartphone. It’s been a recurring dream since Edward Thorndike first put a hungry cat in a puzzle box in 1898 and established the foundational principles of behaviourist learning, which were built on by Sidney Pressey in the 1920s and then BF Skinner with his pigeons.

I’ve just read Audrey Watters’ excellent book Teaching Machines which describes this genealogy in great detail, exploring both the scale of the promise and the ultimate failure. I’d thoroughly recommend it. As she explains, Skinner’s machines might have failed to deliver the educational revolution he imagined but he did demonstrate a principle which modern-day machine-makers are still exploiting.

Namely, a reward delivered on an unpredictable schedule, sometimes after one correct response, sometimes after several, produces behaviour far more persistent than a reward every time. This is the variable-ratio schedule and it is the single most reliable method we have for making a behaviour hard to stop.

It’s the mechanism inside a slot machine, and it’s the same reflex exploited by the ‘pull-to-refresh’ gesture, the drag-and-wait that Twitter made ubiquitous, where sometimes the feed delivers something worth having and sometimes it doesn’t.

Significantly, Skinner didn’t use variable-ratio schedules in his own teaching machines, believing that classroom learning should be errorless and guided by continuous, immediate reinforcement. He felt that unpredictable failure caused anxiety and frustration and, instead, he wanted students to experience continuous success to build confidence and mastery incrementally.

He didn’t show quite the same understanding to his ping-pong playing pigeons and used variable-ratio schedules to condition their behaviour. It was a deliberate choice to keep the two things separate.

Modern learning machines make no such distinction and seek to conflate the two methods; they remove the friction to make the learning feel effortless and they deploy the absorption techniques Skinner reserved for animals. Both become bundled together in the now-ubiquitous notion of engagement.

This fusion has a name. We call it gamification.

I remember reading articles when the video game Call of Duty: Black Ops III was released in 2015 which suggested that it was played, cumulatively, for 75 million hours in the first weekend, a period which far exceeds recorded human history. I don’t know how true those claims really are but I do remember conversations with colleagues and a raft of articles in the educational press all wondering what it would be like if we could leverage some of that engagement for Shakespeare…or algebra…or (fill in the blank according to your particular discipline). That really would be the holy grail of education.

It’s a worthy ambition and not without its merits but it treats engagement as if it is a substance that exists separately from the thing that generated it, a fuel that could be collected in a huge tank by playing Call of Duty and syphoned off illicitly to drive the study of King Lear instead.

To be clear, I am not against the idea of gamification and Kim and Castelli’s meta-analysis does demonstrate positive effects, while also noting that they tend to be short-lived and to diminish as the novelty fades. The reason the gains are shallow and temporary is because gamification tends to weigh heavily on the absorption side of the balance and much less on the friction side. It makes the task stickier without making it harder in the way that produces durable learning. Indeed, by removing friction and making it seem easier, you ensure there is nothing left when the novelty of the scoreboard or the streak has worn off.

None of this is new. Over a decade ago, Rob Coe listed engagement among his “poor proxies for learning“, the things we observe in a classroom and mistake for the learning itself. More recently, David Didau has pushed the idea further, warning that even the correctives become ‘costumes’. Strip out engagement and we simply start worshipping effort and rigour instead, mistaking the appearance of difficulty for the real thing.

He is quite right that desirable difficulties have to be desirable, and that undesirable difficulties are just difficulties. But that leaves the teacher, or the school choosing a piece of software, with some genuinely hard questions. If difficulty is not automatically good, and ease is not automatically bad, how do we tell, looking at a particular tool, whether the friction it contains is the kind that teaches or the kind that merely obstructs?

We need a test.

We can begin with a single question. Thinking about an edtech tool, we can ask:

Does it remove the productive struggle that creates the learning, or work that was never the point of it?

Consider a calculator. In a mechanics lesson, a calculator removes the arithmetic so the student can concentrate on the forces. Doing the sums isn’t the real point but, for many students, it represents significant friction and offloading those sums frees attention for the important thing.

In an arithmetic lesson, the same calculator removes the entire object of the exercise. The tool hasn’t changed. What it does hasn’t changed. What has changed is whether the sums were the friction that constituted the learning or whether they were friction that got in the way of it.

You can’t answer by looking at the tool; you have to consider precisely the purpose you want it to serve.

We can ask a similar question if we are thinking about setting a task for our students:

Does it demand the productive struggle that creates the learning, or work that was never the point of it?

So far so good. Unfortunately, this is the point where things get a bit messy because the tool we all need to be thinking about does not behave like a calculator at all.

A calculator has edges. It does arithmetic and that’s broadly it. So, the question “what work does it remove?” has a stable answer and we only need to ask it once.

Artificial Intelligence has no such edges and that makes things problematic.

The example of a spell-checker illustrates just how things have changed. A ‘classic’ spell-checker had edges; it just compared what you had typed against a stored list of words and put a red squiggly line under the anomalies. It did one job and, if you wanted to know how well a student could spell, you would prevent its use; if you were interested in something else, you could allow it. We all know that a modern spellchecker doesn’t just check spelling but suggests full sentence rewrites, alters tone, generates vocabulary, and completes your thoughts.

What it takes away from the learner is not fixed by what it is; it is decided, exchange by exchange, according to how it is used. The student who asks it to critique the paragraph she has written and the student who asks it to write the paragraph for them are using the identical software to opposite ends. One has offloaded the proof-reading. The other has offloaded the education.

What’s more, no-one deliberately chose to swap the old red-squiggle checker for sentence-rewriting AI. It just happened invisibly, update by update so a school that decided in 2020 to allow spell-checkers in Microsoft Word is now permitting something entirely different without ever revisiting the policy.

So, we can’t apply the same ‘tool test’ to an AI tool, simply because it doesn’t stand still long enough.

We can’t even apply the ‘task test’. In the old days, a task was a fairly fixed thing and you were broadly in control of it. You could plan it to target a particular skill and it would work more or less well in practice. But where the task might have been ours, open to evaluation in the planning stage, the AI exchange is the student’s, happening in a moment when we are unlikely to be present. The object of assessment has moved from something we control to something we can only inspect after the fact, if at all.

So we need a new question. Not “Is AI good for learning?” but:

What did the learner hand over to AI, and was it the thing they were meant to be learning to do?

That question does not have a permanent answer. It has to be asked again every time. Which is inconvenient and a great deal harder than measuring how long a child stays on an app. It is also, I think, the only question that reliably tells the difference between a tool that teaches and a tool that merely does the work.

I am aware that this is a demanding conclusion. I have spent two articles questioning the effectiveness of the measures we tend to reach for. I’ve suggested that engagement is inadmissible, that mere difficulty is easily mistaken for the productive kind, that the tool won’t hold still - and I have replaced them with a question you have to ask for every student, every task, every exchange.

Worse, as Lodge and Loble demonstrate, the answer depends not only on what the student handed over but on what they already knew. A pupil who can judge the quality of what the machine returns is doing something quite different from one who cannot, even when the request is identical. The productive use of AI turns out to be conditioned by the very knowledge the tool is supposed to help build.

None of this fits on a poster, and I am suspicious of anyone who tells you it does.

Which is why my next series will not try. Instead of a rule, I want to apply this question to specific situations and specific subjects and see what it actually demands there. What does the productive struggle look like in a maths problem, a science practical, a history essay, a piece of creative writing? Where, in each, is the line between offloading the friction and offloading the learning? I don’t yet know all the answers. That is rather the point: if there were a simple answer, I would distrust it. The work is difficult and it lies in the particular details, but that is the challenge we have to embrace as teachers.

Which brings me back to Soma. In the novel, Mustapha Mond explains to John the Savage:

“There is always soma to calm your anger, to reconcile you to your enemies, to make you patient and long-suffering. In the past you could only accomplish these things by making a great effort and after years of hard moral training. Now, you gulp down two or three half-gramme tablets, and there you are.”

The disquieting thing about Huxley’s drug was not that it failed. Rather, it worked perfectly, doing just what it was designed to do and everyone was content. It removed the friction, but no-one noticed a problem because nothing appeared to be wrong.

Similarly, an app that removes the struggle from learning does not look like it’s failing either. The child looks absorbed, the scoreboard appears to be mounting, we’re collecting ‘data’. Everything looks good. Like Soma, frictionless learning requires us to notice what is missing in the moment and then to insist on retaining the discomfort that keeps us learning because that is what keeps us human.

Connect With Me

Beyond my writing and tutoring, I work directly with schools, educators, and organisations navigating AI integration. Take a look at my website and please get in touch - I’d love to hear what you’re working on.