The Appearance of Learning
Artificial Intelligence and Educational Inequality

Neil Selwyn’s (2022) article “The future of AI and education: Some cautionary notes” sets out five cautions about artificial intelligence in schools: judge “actually existing” AI and ignore oversold speculation; remember how poorly AI models social contexts and human judgement; attend to its social harms; recognise that claims made for it carry values; and count its environmental cost. He concludes that using AI in education is a political action, never a neutral one, with different effects on different groups.

The article appeals to me because I have lived on both sides of it: in Social Studies teaching, and in creating Te Kete Ako (tekete.co.nz), a library of curriculum resources for teachers in Aotearoa, built with AI. The website is the surface of a much larger AI world model, so as AI capabilities grow the site should keep pace far better than a traditional educational website. The futures lecture suggested re-reading Selwyn’s earlier work “seven years later” to see “how the arguments have moved on (or not)” (Morgan, 2026b). His 2022 article appeared weeks before ChatGPT was released, so I tried the same exercise, drawing also on his 2016 chapter and wider publications.

My first response was to disagree. The AI that exists in 2026 does things that counted as speculation when he wrote, and I work with it daily. In a course reading, Selwyn (2016) repeats Larry Cuban’s verdict that “computer meets classroom, classroom wins”. I am less sure the classroom wins this time. The same chapter warns that “alarm bells should start to ring as soon as anything is presented as being inevitable” (Selwyn, 2016, Chapter 1). My confidence in the technology is a claim of that kind, and reading him made me notice it. His second caution also survives: a more capable model still has to represent a social setting.

I watched the problem in my own classes. A Social Studies class played Blooket and Kahoot quizzes on Te Tiriti o Waitangi several times. Their scores rose each round and they became very quick at answering. When I later asked the same questions aloud, without options, several students who had chosen the right answer in the game could not tell me where Te Tiriti was signed, one of the easier questions in the quizzes. Recognition and recall are different demands, so I cannot say they learned nothing. What troubled me was that the platform’s record looked better than the lesson I found by talking to them. The futures lecture offers the dashboard as the image of our present, where “complexity may exceed our capacity to control it” (Morgan, 2026b). I had the dashboard, and it knew less than a conversation did.

Bishop and Berryman (2009) give me the measure Selwyn’s second caution needs. The Te Kotahitanga effective teaching profile grew from Year 9 and 10 Māori students’ accounts of school, which centred on relationships with teachers. It asks teachers to reject deficit theorising, to accept that they can change what happens for Māori students, and to practise manaakitanga, caring for students as culturally located human beings. An adaptive platform personalises from a student’s history of answers, which our Edu-tech lecture called “a revival of behaviourism” (Morgan, 2026a). When it sends a student easier questions because of earlier errors, it delivers a low expectation automatically. That is a deficit theory with no theorist to challenge. “Personalised” covers responsiveness to data and responsiveness to a person, and Bishop’s teachers were asked for the second.

This shaped Te Kete Ako. I aim the AI at teachers and their preparation, which seems safer and fairer than aiming it at students, and leaves the lesson free to run on discussion, paper and relationships. Selwyn’s fourth caution still lands on me. I am building a locally grounded alternative to imported platforms, attentive to tikanga and mātauranga Māori, on infrastructure owned by a few overseas companies whose values I did not choose. I have not resolved that.

This course began with Bernstein’s (1970) claim that education cannot compensate for society, and Bourdieu and Passeron (1977) suggest why AI is unlikely to be the exception. For them the school reproduces advantage by looking neutral: it treats the cultural capital some children inherit as individual ability, so unequal outcomes appear earned. Give every school the same AI tutor and that appearance grows stronger, because everyone seems to have had the same chance. A well-resourced school can place AI beside expert teachers, a library and small classes, and decline it where it adds nothing. A school under financial pressure may be sold a platform as a cheaper substitute. This is only a prediction. If it is right, the advantage of affluent students will lie less in having more technology than in being free of its indiscriminate use.

Selwyn writes as a critic of the technology, and I will keep building with it. Reading him has changed the question I ask of my own work, from what the system can do to his question, “In whose interests does this work?” (Selwyn, 2016, Chapter 1). The futures lecture asks what kinds of futures education should help to create (Morgan, 2026b). The one I fear is a society that becomes very good at producing the appearance of learning while the conditions for deeper learning are shared less and less equally.