A common response to AI in higher education assumes a clean separation: first teach foundational skills without AI, then build AI-augmented performance on top. The foundational skills are supposed to be specifiable independently of the technology. Call this the Robinson Crusoe picture of learning. There is a core of competences the individual must have on their own, and technology enters only after those are in place.

This picture is mistaken, and recognizing why matters for how universities respond to AI. Base-level skills have always been defined relative to available cognitive tools. What counts as adequate mathematical competence depends on whether calculators are available; what counts as adequate research skill depends on the information retrieval systems you can use. Each major cognitive tool reconfigures what individuals need to be able to do unaided.

AI forces a more radical version of this point. Previous tools augmented specific capacities while leaving the connection between a person's cognitive ability and their output largely intact. AI severs that connection. A student with modest understanding can now produce work that is hard to distinguish from expert output. The evaluative concepts that education relies on (intelligence, understanding, skill, competence) were built for a world where output was a reliable signal of individual capacity. That world is disappearing.

I argue that the right response is to build new evaluative categories suited to a world where cognition is always embedded in technological systems. I propose epistemic agency as a central candidate: the capacity to function as the responsible human component in a human-AI cognitive system. This requires calibrating trust in AI outputs and recognizing where one's own understanding falls short, so that the human retains genuine responsibility for the joint output even when much of the cognitive work has been offloaded.  

The talk outlines what replacing some of our existing evaluative categories would require and what that means in practice for curriculum and assessment design.

Herman Cappelen is Chair Professor of Philosophy at the University of Hong Kong. He is the founder and Director of the AI & Humanity Lab at HKU, Director of the MA programme in AI, Ethics and Society, and Editor-in-Chief of Inquiry: An Interdisciplinary Journal of Philosophy.

Before moving to Hong Kong, he held positions at the Universities of Oslo, St Andrews, and Oxford. He received his BA in Philosophy, Politics, and Economics from Balliol College, Oxford, and his PhD in Philosophy from the University of California, Berkeley.

His research focuses on the philosophy of AI, Conceptual Engineering, and the connections between the two. His recent books in the philosophy of AI include Making AI Intelligible: Philosophical Foundations (with Josh Dever, OUP, 2021), the first monograph on metasemantics and large language models, and the forthcoming Going Whole Hog: A Philosophical Defense of AI Cognition (with Dever, OUP). He is also co-editor, with Rachel Sterken, of the forthcoming volume Communicating with AI: Philosophical Perspectives (OUP). His papers on AI cover topics including existential risk, AI introspection, and externalist approaches to AI content.

His work on Conceptual Engineering includes Fixing Language: An Essay on Conceptual Engineering (OUP, 2018), the first monograph on the topic, and The Concept of Democracy: An Essay on Conceptual Amelioration and Abandonment (OUP, 2023). Earlier books include Insensitive Semantics (with Ernest Lepore, 2005), Philosophy without Intuitions (2012), and Relativism and Monadic Truth (with John Hawthorne, 2009).

He is an elected fellow of the Norwegian Academy of Science and Letters, a permanent member of the Institut International de Philosophie, and an elected member of the Academia Europaea.