What Do We Want to Remain Capable Of?

Learning scientists Robert and Elizabeth Bjork call these “desirable difficulties”: things like recalling an answer from memory, spacing your practice out, or trying a problem before you see the solution. They make learning feel harder now, but they make it last. When AI removes that effort, it can quietly remove the learning the effort produced. A 2026 review of 67 studies names the same mechanism, epistemic friction: without it, “AI-generated fluency can bypass the reflective struggle central to deep learning.”

In a Harvard physics study, a purpose-built AI tutor designed with proper scaffolding beat in-person active learning by 0.73 to 1.3 standard deviations, which the authors describe as a large effect. In the Türkiye study, the standard ChatGPT left students worse off once it was removed, while a guardrailed tutor version essentially avoided that loss. The same model could harm learning or spare it, depending on how the tool was designed.

In a controlled study in a university creative-coding course, students received identical AI-generated feedback on their work. Those told it came from a human teaching assistant ran their code more, wrote more code, and spent more time on later work. They rated the feedback equally helpful either way. The effect on effort was large (d = 0.88 to 1.56). The content was the same.

Education does three things at once: it builds knowledge and skills (qualification), it helps you find your place among others (socialisation), and it helps you become someone who thinks independently and takes responsibility (subjectification). AI tools mostly reach the first. They rarely, if ever, address the other two, and those are where a teacher does their deepest work.

An LLM answers your question. Faster work, less learning.

An AI tutor asks questions back, no matter what you actually need. Often frustration and drop-out.

An AI learning companion (Dr Philippa Hardman calls it a “study mate”) remembers where you got stuck and pushes you towards the thinking you avoid. Capability that lasts.

This map keeps meeting three stakes: what AI does to our thinking, to how we live and learn with others, and to the world it draws on. They are usually studied apart. Learning scientists measure thinking, ethicists debate fairness and power, engineers count energy. The declarations that try to hold them together, from the IEEE’s design guidelines to the Montreal and Toronto declarations, tend to do it around a single aim: human well-being.

In 2019 a group of scholars, artists, technologists and knowledge keepers, most of them Indigenous, from Aotearoa, Australia, North America and the Pacific answered those declarations with a position paper of their own, published with the Canadian Institute for Advanced Research and edited by Jason Edward Lewis (Cherokee, Hawaiian and Samoan), who holds a research chair in computational media at Concordia University in Montreal. Their objection is precise: “none of these efforts challenge the fundamental anthropocentrism of Western science and technology.” Their alternative is relational. Instead of asking only how a technology serves us, they ask what relationships it creates, and what each side owes. In their words, “while the developers might assume they are building a product or tool, they are actually building a relationship to which they should attend.”

Read this way, the three stakes become three relationships. With what we know: the historian Noelani Arista (Kanaka Maoli) asks whether “computer memory” will “replace experts and elders as repositories of knowledge”, and calls for institutions that train knowledge keepers “fluent in language and trained in computer science.” With others: for Caroline Running Wolf and Noelani Arista, research and applied projects need to be built “collaboratively with, not on behalf of, and certainly not without the community”. More broadly, the paper asks communities to build their own values into these technologies, since “the alternative is to have our worlds designed for us.” With the planet: the artist Suzanne Kite (Oglála Lakȟóta) asks “What is being offered to the Earth when we extract these mined materials?”, the question the previous section began to answer in kilowatt hours.

This is a lens, not evidence, and it asks for care. The paper does not claim a unified voice (“Indigenous ways of knowing are rooted in distinct, sovereign territories”), and Suzanne Kite writes that her essay “does not attempt to speak for all Lakota”. The paper does not forbid others from learning from it, but many of its participants “expressed concern about issues of appropriation and misuse of traditional knowledge”, so it is credited here as a way of seeing the question, not borrowed as a method. It was written in January 2020, before ChatGPT, and gives a central place to systems that communities control or build themselves. And as a position paper, it opens a question rather than settling one.

The tool: the guardrailed tutor and the learning companion are instructional design built into software: scaffolding, answers withheld, help that fades as you grow.

The classroom: co-design tools with teachers rather than deploying them on teachers, and set tasks that make learners compare, justify and revise what the AI produces. In the 67-study review, that scaffolding is what separated gains from cognitive offloading.

The system: AI only helps where the conditions are ready. As the WEF puts it, “learning outcomes will not be determined by technology itself, but by the conditions in which it is deployed”, and isolated fixes across policy, pedagogy and technology are unlikely to be sufficient.

In August 2026, the system layer got its first institution-scale test case. MIT’s committee on AI in teaching and learning refused patching (“this is not a moment for patches and duct tape”) and called for all three moves in concert: AI-aware redesign of every subject, an in-person social component in every subject, and permanent adaptation structures (a standing committee, per-school AI leads, a pilot fund), on a platform deliberately tied to no single AI vendor. One survey number shows what is at stake: MIT undergraduates report feeling more replaceable than capable (40% versus 34%).

Susskind adds the one move here with a precedent. In 1982 a British government report by Wilfred Cockcroft answered the electronic calculator not by banning it but by splitting mathematics in two: part of the time learning to work with a calculator, the rest learning to cope without one, and, the decisive part, examining both. That split is now the standard almost everywhere. He proposes the same structure for AI, in every subject, and calls it teach both, test both. His reason for putting the weight on assessment rather than detection is the part worth keeping: a teacher cannot know whether a student used AI alone in their bedroom, but nobody forgets sitting an exam they have only half prepared for.