Pedagogy, Not Technology

In short:

• Goal → pedagogy → tool: treat AI as the means, not the end.

• Three visions—Conservative, Baseline, AI‑Forward—show how schools can benefit today, soon, and one‑day.

• By the Varian Rule, learning experiences that once belonged to the 1 % will reach every classroom.

The scarce resource in classrooms of tomorrow is teachers' understanding of their students' needs.

Prompt packs, chatbots, and dashboards sell the fantasy that tools alone transform learning. In a lot of the social media discourse I see about AI for teachers, the tool becomes the destination, and pedagogy retrofits around its quirks. But the correct steps to think about AI in education is surely to 1) start by defining learning goals or problems to solve based on teachers’ understanding about their students, 2) choosing the instructional design, then 3) inviting AI in as helper.

As if that wasn’t enough work already, there should also be a step zero to agree on how much we want to change teaching and learning to accommodate AI. Below, I’m throwing out three visions to set the stage for something better, on how that change could play out at increasing levels of ambition.

• Relieve teachers’ burden
  An Apr ‘25 survey shows GenAI users save at least 30 min/day because they spend less effort drafting and more on verification. More experienced users can save even more time while delivering better results. Have teachers picked all the low hanging fruits already?

• Personalise learning
  38% of teachers in the UK used AI to create lesson content, model answers, and for lesson planning. Are they taking advantage of the speed and ease of AI to generate differentiated sets of materials for each lesson, or using more frequent pop quizzes to surface misconceptions before they fossilise?

• Cognitive pivot: verification & integration
  A Microsoft study finds that “GenAI shifts the nature of critical thinking toward information verification, response integration, and task stewardship.” In practice that means teaching students a three‑step fact‑check protocol (source, corroborate, reflect) every time an AI draft appears.

• Differentiate AI‑assisted vs. independent work
  Anthropic’s million‑chat analysis shows student usage patterns ranging from quick answer‑grabs to iterative co‑creation (an echo of Vygotsky’s social‑constructivist “more capable peer”, now in silicon). That means a two-tiered assessment framework to test solo competence where it matters, while new rubrics might score AI‑assisted work for prompt quality and iterative refinement, not just final prose.

• Toward universal Two‑Sigma gains
  The Two‑Sigma problem is Bloom’s 1984 finding that 1‑to‑1 tutoring moves the average student two standard deviations above the conventional class mean, but cost made it elusive. A World Bank pilot let Nigerian pupils use GPT‑4 twice a week and delivered roughly two years of learning gains in just six weeks. Technology now offers a plausible path to scale 1‑to‑1 tutoring for all.

• Teachers’ new edge — curating the environment
  When pacing, practice, and instant feedback are automated, teachers’ irreplaceable value lies in crafting supportive, inquiry‑rich cultures where students negotiate meaning, ethics and identity—spaces no algorithm can reproduce. This is the long‑promised guide‑on‑the‑side role that we have discussed for decades, except now we actually have the tools to make it happen.

Say No to AI for the Sake of AI

Whatever vision makes sense to you, begin with what only you know best: your students’ aspirations, barriers and quirks. Some classes will thrive on multimedia storytelling, others on adaptive practice, others on rigorous debate. Do students need scaffolds, drills, or space to wrestle with ambiguity? These are the questions we need to ask to put pedagogy in the driver’s seat.

AI can already draft, translate, storyboard, and simulate, and soon it will do more we haven’t imagined. In that world, insight into what learners truly need is the scarce resource that only schools and teachers can provide.