Reflections on the AI in Education Summit

Andromeda, Juan Antonio Escalante, C17th. From Wikimedia Commons

I was lucky enough to be a delegate at Timo Hannay’s excellent AI in Education Summit last Tuesday.

The discourse around Generative AI (GenAI) in education is turbulent, with a wide range of hopes, doubts and ideas tumbling under one heading. In the background of every conversation is the sense that our state education system is passive and helpless, like Andromeda chained to a rock waiting for the monster—and perhaps also the hero.

I came away wanting to break up that discourse a bit, to pull generic questions of what could/can/will we do about GenAI into parts addressing the distinct risks and opportunities this technology opens up, without every time pulling on a thread that ends with us chained helplessly to that rock. AI anxiety is real, and it’s justified, but it’s also difficult to make progress when it’s in operation. I’m acknowledging it in the hope of setting it aside.

Who is this for?

One obvious way to split those anxious general questions is along the lines of different phases of education. As Tom Chatfield pointed out, GenAI already affords those with the necessary skills and willpower an extraordinary interlocutor for personalised learning, with ad-hoc Socratic dialogue, instant examples, and endless patience—a “cognitive copilot”. But this kind of self-directed work is closer to undergraduate study than the typical primary classroom. The broad heading of “education” isn’t much help here.

There is the tantalising possibility, which Tom’s brilliant session touched on, that GenAI itself will help us to push such activities into children’s learning practice earlier on. Perhaps this tech will find a place in the classroom. But we need to treat the needs of autonomous adult learners as separate from the needs of children, even if the responses end up overlapping.

Reform opportunities

Whilst much commercial energy is focussed on personalised learning experiences like those “cognitive copilots”—and those conversations lead into interesting corners, like our relationships with “our” AIs, solipsism, and unreliable “teachers”—curriculum reform is bubbling under everything.

For Conrad Wolfram, the crisis is an opportunity to reinforce existing arguments for reforming STEM teaching to acknowledge the power and utility of computers. For Tom Chatfield, the indeterminacy of GenAI creates a good imperative to instil critical thinking skills early. And of course, because everything is mixed up and we’re all Andromeda, the discourse around that compelling argument typically leans towards inviting GenAI to be the cure for its own disease: using GenAI to teach us how to handle GenAI.

Wishing to separate these two, I wondered what an AI-enabling curriculum which had no AI in it at all would look like. There’s an interesting thought experiment (which might not be an experiment for long!) in asking what skills we might need in a world where we are sometimes discouraged from thinking critically about the outputs of GenAI. In any case, we need to decide what we’re going to put on the curriculum and how we’re going to assess it—and we need to not get distracted by GenAI itself while we do it. It’s clear that the changes will need to be quite radical.

How will we know when we’ve got there?

We need to talk about feedback and measurement. We’re entering a phase where technology (privately-owned, elective, constantly changing) suddenly has the power to drive schools (state-owned, mandatory, hard to change) in unexpected, unofficial directions, without understanding the trade-offs of one reaction or another. For example, GenAI’s power to support cheating on homework is creating direct and immediate pressure on teachers. How should we intervene?

A fascinating conversation tackled the evidence requirements for GenAI-related interventions into education, with voices speaking in favour of traditional randomised controlled trials or shorter, faster feedback loops, and others simply observing that the horse had bolted and we were now in a natural experiment. But even as we started talking about the subject, we were already mixing up, for instance, the use of GenAI for teaching (an opportunity) with the use of GenAI for cheating (a risk), and of course the use of GenAI as the remedy.

In future, to get a handle on the multiple effects of GenAI and to have accountability, we need public, open feedback loops showing the effects of each of those things, not “AI” as a whole. And we need to split and name these different strands if we’re going to achieve that—but the language doesn’t seem to be there yet.

Dependency

The last separation I want to propose is between different kinds of dependency: individuals depending on machines for cognition; society as a whole doing the same; and the technical dependency that comes with private model ownership. The first is strongly emotive and quite novel, the last is not. But it matters.

The Andromeda mindset treats LLMs as a single indivisible force. In fact, they are artefacts: highly various, fine-tuneable, open-sourceable, adaptable and (in due course, we hope) inspectable. As we limber up to run large sections of our economies, let alone our schools and universities, on these systems, we ought to hold on to these levers. We can, if we have the technical know-how. If we don’t, we’re like any other enterprise software customer, locked into years-long contracts, paying endless upgrade fees, waiting for fixes and features.

In that context the risk is not runaway revolutionary tech, it’s boring old vendor lock-in, with an occasional OpenAI-style corporate implosion thrown in. That the tech is sexy doesn’t make that aspect go away.

Conclusion

I write all the above as a novice: I’m not an educationalist, I’m a software developer. But I’m thrilled that this debate is open and that we can attend conferences about it—thanks Timo! Thanks also to Tom Chatfield and Conrad Wolfram who gave permission to use their names under Chatham House rules.