After the Proxy War
Winning the peace
In her 1959 book, An Introduction to Wittgenstein’s Tractacus, Elizabeth Anscombe relates a conversation she had with Wittgenstein:
He once greeted me with the question: ‘Why do people say that it was natural to think that the sun went round the earth rather than that the earth turned on its axis?’
I replied: ‘I suppose, because it looked as if the sun went round the earth.’
‘Well,’ he asked, ‘what would it have looked like if it had looked as if the earth turned on its axis?
Anscombe is talking about Wittgenstein’s later approach to dissolving philosophical problems by examining the linguistic and conceptual frameworks that generate them. Wittgenstein wasn’t trying to criticise the medieval astronomers, many of whom were highly sophisticated mathematicians. Rather, he was interested in why a particular way of seeing things feels so natural that alternatives become literally unimaginable.
The geocentric model wasn’t just a belief that people had; it was the perceptual frame through which they saw everything. And that framework continued to provide evidence to support itself long after Copernicus had proposed the heliocentric model. People didn’t suddenly see that they had been wrong all along, as if in a moment of revelation; Copernicus initiated a shift that was long, slow and hotly contested.
Not all our perceptual frameworks are as significant or as all-encompassing as the geocentric model of the universe but they can be just as persistent. Indeed, they are invisible precisely because we are inside them; we don’t have a vantage point outside of ourselves from which to see. As Wittgenstein put it:
A picture held us captive. And we could not get outside it, for it lay in our language and language seemed to repeat it to us inexorably.
(Philosophical Investigations (1953)
As teachers, we find it just as hard to imagine alternative ways of seeing. I have set thousands of essays to students over the years in the belief that I was gaining a genuine insight into their understanding, so the idea that the essay form might be a proxy that AI is quietly exposing is a challenging and uncomfortable one.
Bertolt Brecht, the great twentieth century German playwright and theatre director, understood the problem Wittgenstein described and proposed a mechanism to help us to see things afresh.
Given the German fondness for hugely long composite nouns, he called this Verfremdungseffekt. Given that we struggle with them in English, it’s usually called V-Effekt or, even more commonly, ‘alienation effect’, although I find that translation problematic because it carries psychological baggage which runs against what Brecht intended. Both his word and the Russian concept of ostranenie which inspired him are more to do with ‘strangeness’.
He described the V-Effekt stripping events of their self-evident, natural, familiar quality and instead making them strange or surprising or requiring explanation. For example, where audiences were used to enjoying the suspense of a play and wondering what would happen next, Brecht would sometimes reveal the outcome on placards at the side of the stage. The idea was to free the audience from thinking about what was going to happen so they could think instead about the why and how.
Brecht wanted to reveal the mechanism of the theatre, all the backstage space, the rails of costumes and so on that are normally hidden from the audience. His aim was to make the structural framework visible and unavoidable.
AI is doing precisely the same thing in the world of education. All of the familiar instruments of teaching and assessment that we have relied on for so long suddenly start to look a bit strange and, perhaps, not quite as reliable as we had assumed they were.
We can choose, like many contemporaries of Copernicus or Galileo, not to look. Or, we can move forward, hoping that the new frameworks we create will be even more useful to us.
In his 2013 Ted Talk, Sir Ken Robinson uses the metaphor of Death Valley to suggest that, at every level of education, the fundamental role of educators is to create the conditions, a ‘climate of possibility’ where people will ‘achieve things that you completely did not anticipate and couldn’t have expected.’
So, the vital question to ask is: What might those conditions look like?
When Jørn Utzon was invited back into a consultancy relationship with the Sydney Opera House in 1999, thirty years after he had left the project under some duress, never to return to see the building opened to the public, he wrote a remarkable document setting out the fundamental principles that underpinned his design. He understood, and maybe even relished, the fact that his building would change over time but he wanted to establish a distinction between the fundamental character, the experiential intent which should be preserved and the elements that should evolve through use and technological development.
It is in that spirit that I want to offer not a blueprint for education in the age of AI but a set of principles against which we can assess the choices we are making, so that what we build next can remain true to the essential purpose of education: to help people to learn. None of what follows is certain. It is conditional on choices - institutional, pedagogical, political - that have not yet been made. All of it, that is, except the last one.
Proxies in education are inevitable - but they must remain honest. A student’s learning will always be largely invisible, existing in their head and we have to have a mechanism, a proxy to measure it somehow. In this series of articles, I’ve tried to highlight those proxies that have become invisible to us, those we have ceased to question because they feel inevitable or natural. It is important we continue to interrogate the gap between the measure and the thing we want to measure and that we make adjustments when the connection drifts. So, remember Goodhart’s Law: When a measure becomes a target, it ceases to be a good measure.
Cognitive struggle must be preserved by design. Not all friction is productive and not all ease is harmful. We need to be clear-sighted about the purpose and function of the activities we design for our students and understand when and where the ‘desirable difficulty’ needs to be located. I’d suggest it should be in the thinking, the evaluating, the forming and defending of a position, not in the mechanical retrieval that AI can perform more efficiently than any student. If AI can do the task without thinking, the task probably wasn’t targeting thinking.
Student thinking must precede AI input. As teachers, we understand that the sequence is important and AI makes it even more so. A student who forms a position and then tests it against AI is doing something categorically different from one who asks AI to form the position for them. We need to help our students to see AI more as a challenger and less as a generator. Not easy, I grant.
Strategic AI use must be explicitly taught. It is becoming clear that the way students use AI leads to radically different outcomes. Students who simply hit the ‘easy button’ to fulfil the task set by the teacher are likely to learn much less than those who use AI strategically. But strategic use of AI is a transferable skill that has to be taught explicitly and consistently; it is not an attitude that students absorb simply by proximity to the technology. If all students are taught these skills, there is a small chance that AI will have the equalising impact many of us hope for. If the knowledge is left to distribute itself, it is likely to flow along existing lines of advantage. The Matthew Effect, the well-documented tendency for educational advantage to compound, will be turbocharged.
Assessment must focus on process, not just product. The prompt history, the question trail, the source annotation, the dialogic exchange are all attempts to narrow the gap between the measure and the thing being measured. The principle is that a student’s thought process is more revealing than what they produce at the end of it, and assessment systems need to be designed to see the former rather than rewarding only the latter.
Feedback should be immediate, specific and responsive. I’m aware that in this series, I have argued in favour of AI feedback and have been sceptical about AI marking platforms. That apparent tension can be resolved by thinking about the precise purpose of the task. Students need immediate feedback to make effective progress with an activity but it is not always possible for the teacher to give that, either because there are thirty other students demanding the same or because the student is at home. AI can usefully plug that gap but I don’t believe it can replace the nuanced, targeted feedback of an experienced professional. The two are fulfilling different functions and they can and should co-exist.
Metacognitive awareness must be cultivated deliberately. If students are to engage critically with AI output rather than accept it wholesale, they need to understand something about how their own thinking works. This is the condition that strategic AI use most directly strengthens but, again, thinking about thinking must be taught and not assumed.
Institutional conditions must actively support scrutiny. Teachers cannot interrogate the frameworks they are inside without leadership that creates space for that discomfort. The principle is that schools and systems need to create genuine permission for teachers to experiment, to acknowledge what isn’t working, and to redesign practice in response. That is extremely difficult given the institutional orientation towards the exam as the centre of the universe. That doesn’t make it any less important.
The Unconditional Principle: The human relationship must remain at the centre. This is the one principle that is not conditional on how AI is used, what the institution permits, or what the assessment system rewards. Learning is not a hidden inner event produced by sufficiently good delivery. It is a disposition that develops in a whole person, over time, through struggle, feedback and the accumulated relationship between a particular student and a particular teacher. AI can perform the task-sense functions of teaching with increasing fluency. It cannot know the student, respond to what is unspoken, or create the conditions of trust and challenge within which genuine understanding becomes possible. That is not a temporary limitation waiting to be engineered away.
It is what teaching actually is.
If we view AI as a V-Effekt encouraging us to challenge and question the frameworks that have become so familiar that they have become invisible, then maybe we will remember that.