Giving Up the Ghost
Why the failure of the "Teacher-Machine" marks the beginning of AI’s real life in the classroom.
In 2019, Amanda Spielman, then the Head of Ofsted, the body responsible for school inspections in England said:
“We know that a data-heavy culture has led to perverse consequences, consequences that are actively detrimental to the substance of education. We’ve seen curriculums being narrowed and a culture of teaching to the test sometimes trumping real learning. I understand why these practices have emerged, and I acknowledge that we have played a part.”
This was big news at the time.
It shouldn’t have been. Matthew Arnold, the poet, cultural critic and school inspector said something very similar 157 years earlier.
Writing in 1862 about Robert Lowe’s Revised Code, Arnold observed that inspectors were reduced to hastily glancing around the school before retreating to the logbooks
“as if there might not be in a school most grave matters needing inspection and correction“
while teachers, meanwhile, were led to think
“not about teaching their subject but about managing to meet targets.”
This raises an obvious question. If a thinker as influential as Matthew Arnold could see it so clearly in 1862, how is it possible that we are still here?
It’s not just that no-one listened to him. Rather, the problem he identified is structural, built into the very foundations of how we think about teaching and, more particularly, what a lesson is and what it is for.
To understand the origins of this structural problem, we have to go back a bit further to the seventeenth century and Descartes. In his ‘Meditations on First Philosophy’, he argued that reality comprises two fundamentally distinct kinds of substance:
Res Extensa (Extended Substance) - the body and the physical world. Its essence is extension, taking up space. It has physical dimensions that can be measured, moved and divided into smaller pieces.
Res Cogitans (Thinking Substance) - the mind. Its essence is thought and it takes up no space, it has no shape and it cannot be divided into parts. It is private and invisible.
For Descartes, the body was a machine which was directed by the mind. Although he went on to qualify this distinction, arguing that the two are so intimately ‘intermingled’ in the human body as to form a unity, the message that had such a profound influence on western thought was to do with the separability.
By establishing the separability of the mind, Descartes defined knowledge as a purely internal, non-physical property. Once the mind is viewed as a separate ‘container’ for truth, the body’s role in learning is relegated to that of a mere bystander. The pedagogical consequence of this is a transmission model. Because knowledge is no longer seen as something we do with our bodies in the world, it becomes something that is simply ‘moved’ from the teacher’s mental container to the student’s.
Over time, the vehicle for that transmission became the lesson and it was gradually solidified into bricks and timetables. There were two key staging posts in this process. First was Robert Lowe’s Revised Code in 1862, the one Matthew Arnold reacted against. Lowe believed fervently that social policy should be measurable, calculable and efficient. Recommending his policy to the House of Commons, Lowe reassured sceptical MPs that:
“if it is not cheap it shall be efficient; if it is not efficient it shall be cheap.”
Fundamental to Lowe’s policy was the idea of payment by results. He divided the curriculum into “Standards”, creating a linear, mechanical ladder of learning that every child had to climb at the same pace.Every year Her Majesty’s Inspectors visited each school to test pupils’ progress against these standards. Teachers’ salaries depended directly on the size of the grant their school received, which depended directly on how many pupils passed the annual inspection. It’s not hard to imagine the effect that might have on one’s teaching practice; Matthew Arnold was certainly in no doubt.
The second key event was the Forster Act of 1870 which effectively scaled up Lowe’s policy and created hundreds of new Board Schools, all of them built on the assumption that education was a measurable commodity. The school system became a machine and teachers became ‘instructors’ transmitting information as efficiently as possible from one ‘separable’ mind to another under the watchful eye of a data-collecting inspector.
Payment by results was abolished in 1897 after thirty-five years, but its core tenets remained embedded in the system: the belief that inspectable performance is the substance of education, that accountability demands a measurable unit, and that the lesson exists primarily to produce a grade.
Ofsted, established in 1992, reinforced the structure. Essentially, an inspector visited, observed the lesson and graded what they saw. One hundred and thirty years separate Robert Lowe from Chris Woodhead, its first Chief Inspector.
In 1949, the philosopher Gilbert Ryle published The Concept of Mind in which he challenged the Cartesian picture, as it had come to be understood. While Descartes’ logic might have been sound, Ryle suggested he had made what he called a category error by depicting the mind as something separate from the body.
Ryle illustrated his notion of a category error using the story of a visitor to Oxford University. After a tour of the various colleges, libraries and laboratories, the visitor asked,
“I see the buildings and the scholars, but where is the university?”
The error was to assume that the university was another building, something in the same category as ‘library’ whereas, in reality, the university is the collection or organisation of those parts.
We could apply that to the world of education. A teacher might look at a student’s test scores and see the data but still ask’ “Where is the learning?” It would be a category error to assume that learning is a separate, measurable object in the same category as a test result, rather than a complex, ongoing process of change within the student.
Ryle used the memorable phrase ‘the ghost in the machine’ as a derogatory label to expose what he saw as the absurdity of Cartesian duality. If the body is a purely physical machine governed by mechanical laws and the mind is a purely spiritual, non-extended entity, then the mind is essentially a ghost, an invisible operator sitting inside a mechanical suit, pulling levers it cannot touch.
He argued that the ‘ghost’ only exists because of a category error based on the language Descartes used. He treated the mind as if it were a substance, a noun, just like the body. For Ryle, the mind isn’t a part of the person; it is a description of how the person functions.
Ryle went on to observe that many verbs look grammatically similar on the surface but work in fundamentally different logical ways. Some verbs describe ‘tasks’ or activities, things you can be continuously and currently doing. Others describe ‘achievement’, things that are defined by their result. He used the examples of looking and seeing, where ‘looking’ is a task verb and ‘seeing’ is an achievement verb. The one cannot be substituted for the other without a category error. To say “she was seeing for twenty minutes” sounds odd in a way that “she was looking for twenty minutes” does not.
The grammar of “seeing” doesn’t permit continuous activity in the same way “looking” does, because “seeing” is an achievement - it either happens or it doesn’t. Treating them as interchangeable is the error. We should note, by the way, that by using the term ‘achievement’ we are not talking about success or failure in an ordinary sense. The point is logical, not moral.
Sir Ken Robinson brilliantly picked up Ryle’s point in his 2013 Ted Talk and suggested ‘dieting’ is another example where you can be engaged in the activity of something, but not really achieving it.
“There he is. He’s dieting. Is he losing any weight? Not really.”
He went on to suggest:
“Teaching is a word like that. You can say, “There’s Deborah, she’s in room 34, she’s teaching.” But if nobody’s learning anything, she may be engaged in the task of teaching but not actually fulfilling it.”
These examples are interesting because, properly conceived, they are achievement words but they masquerade as task words. Institutionally, we treat teaching as a task, something you can be doing continuously for the 50 minutes scheduled on your timetable. It is observable and therefore it is gradable. Instead, we should define it by its outcome. We have systematically confused the activity with the achievement. The lesson is the institutional form that confusion has taken.
Sir Ken was clear:
“Teaching, properly conceived, is not a delivery system. You know, you’re not there just to pass on received information. Great teachers do that, but what great teachers also do is mentor, stimulate, provoke, engage. You see, in the end, education is about learning. If there’s no learning going on, there’s no education going on.”
The achievement sense of the verb requires a human relationship at its heart because the achievement, the learning realised in a particular person, only becomes possible in the conditions that a human relationship creates.
John Hattie’s findings in Visible Learning for Teachers: Maximising Impact on Learning support this. The highest-impact factors on student learning - feedback, teacher-student relationship, formative evaluation - are all relational and dispositional. They happen around lessons, between lessons and across time. They fluctuate and shift. The lesson as a unit measures almost none of what actually drives learning.
It is clear that the lesson has become a proxy for learning and the institutional mistake we make with it is a category error in Ryle’s sense.
The noun ‘lesson’ behaves grammatically like a container. You can be ‘in’ a lesson, it can ‘contain content’, an inspector can ‘observe’ a lesson and grade what they see. This grammar invites us to see a lesson as the location of learning.
But learning is not a bounded event which fits neatly into a fifty-minute slot on a timetable. It is a disposition that develops unpredictably across time, through struggle, feedback and the developing relationship between a teacher and a student.
Those of you who have followed me through this series looking at some of the assumptions that underpin our pedagogical practice will not be surprised to hear me suggest that AI has not created this proxy but it has exposed it vividly. It has done so by automating many of the task-sense functions of teaching. It can now explain, demonstrate, create retrieval practice, even give adaptive feedback and it can already do those things pretty well.
The temptation for some is to let EdTech get on with it. After all, it can function at a much larger scale and, so the logic goes, have a much bigger impact than any teacher. That was the energy behind Sal Khan’s excited claims in 2023 for his educational chatbot, Khanmigo. You probably don’t need me to tell you that Khanmigo has rather failed to live up to expectations and it seems it is being quietly dropped, despite the enormous resources lavished on it by Khan Academy and Microsoft.
We need to be clear about why it has failed.
The model for Khanmigo and almost all EdTech tools rests on three assumptions, each inherited directly from the Cartesian transmission model.
The first assumption is that knowledge is a substance that can be packaged and moved, from curriculum to EdTech platform to students’ minds. This is to treat knowledge as a property of res cogitans, a mental substance that can be extracted and transmitted.
Second, a teacher’s role is essentially that of a delivery mechanism, so a better delivery mechanism is a better teacher. This is to treat the task-sense of the verb as if it is the whole of teaching.
Third, learning is an internal event resulting from good delivery. If the delivery is personalised, adaptive and infinitely patient, precisely the claims made on behalf of Khanmigo, then optimal learning should follow.
As Ryle pointed out, the Cartesian model implies that there is a ghost in the machine, a hidden inner substance where understanding lives. If understanding is an inner mental event produced by sufficiently good delivery, then a better delivery mechanism should produce more of it. This is precisely the ghost the EdTech companies have been searching for.
Significantly, when EdTech evangelists talk of a ghost in the machine, which they do a lot, they tend to use the term in a way that has shifted away from Ryle’s original coinage. Either it refers to the possibility that consciousness or genuine understanding might emerge from a sufficiently complex AI machine. Or, it refers to what AI currently lacks, the spark of consciousness which separates genuine understanding from sophisticated pattern-matching. In both usages, the ghost is treated as a real thing, either potentially present or regrettably absent.
But that is the opposite of what Ryle meant. For him, the ghost does not exist. It was an illusion conjured by a category error about the way language works.
To consider the possibility that an AI machine has an internal ghost is merely to repeat the original error. The EdTech companies have done that because they inherited, apparently unquestioningly, an education model that had the error hard-baked within it. They have spent billions building increasingly sophisticated machines to house something that cannot be housed anywhere, precisely because it does not exist.
The failure of Khanmigo and similar EdTech models doesn’t just tell us that the technology was inadequate. It reveals something much more important. It reveals what learning actually requires.
If a machine doesn’t work as you hoped, you gradually strip out everything the machine can do well - in this case, the delivery, the explanation, retrieval practice, the adaptive feedback - and what you are left with is the thing that matters. The failure is diagnostic.
Khanmigo might have failed because it was chasing a ghost but it reveals where learning actually lives.
Learning is not a hidden, inner event resulting inevitably from good delivery. It is a disposition. It is something that develops in a whole person, over time, through struggle, challenging feedback and the accumulated relationship between a particular student and a teacher. It is not mystical and it is not a ghost. It is visible, relational and entirely human.
If we are clear about what learning and teaching actually mean and we understand that they are achievement verbs, then we have a chance of understanding what AI is actually for in education.
AI can perform task-sense functions fluently and rapidly. That is genuinely useful, not because it replaces teachers but because it frees them to do the things only a teacher can do. Know the student. Respond to what is unspoken. Create the conditions of trust and challenge in which genuine learning becomes possible.
Ken Robinson is precisely right when he says:
“Teachers are the lifeblood of schools.”
The achievement sense of teaching is not automatable because it was never mechanical. For 160 years, from Lowe’s Revised Code through the creation of Ofsted, the institutional framework has treated the task as if it were the achievement and the lesson as if it were the location of learning.
By performing the task fluently but failing to produce the achievement, AI has made the category error impossible to ignore.
The lesson was never the learning. It was part of the condition for it to happen. And the condition that matters most is the one no machine can provide. A human being who knows you, challenges you and stays with you.
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.