27 March 2026 · LinkedIn
A thousand high school students in Taipei sat down to learn Python this year. All of them had access to an AI tutor. Half were given a fixed sequence of practice problems, the standard escalating difficulty. The other half had something different: an AI system that watched how they coded, inferred what they actually understood, and quietly rearranged the path forward in real time.
The students on the personalised track scored 0.15 standard deviations higher on the final certification exam. That exam was taken without AI assistance. The gain is equivalent to several months of additional schooling.
They didn't work harder. They stayed engaged long enough for it to matter.
This is a Wharton-led randomised controlled trial across ten schools over five months. Not a prototype. Not a pitch deck. Field evidence at scale.
Consider what sits behind that result. The AI wasn't following a script. It was using reinforcement learning guided by a large language model to extract meaningful signals from the students' own interactions: their code edits, their hesitations, the patterns that reveal comprehension or confusion far more reliably than any self-reported survey could.
A tutor that doesn't just teach, but learns how to teach. For each student individually.
Now widen the lens. Thirty children still sit in rows in most classrooms worldwide, facing a single teacher delivering a single lesson designed for one imaginary average learner. A system architected during the Industrial Revolution. Smartboards replaced blackboards, laptops replaced notebooks: the structure never changed.
The Taipei study makes that structure harder to defend. When an AI can dynamically match difficulty to demonstrated mastery, the rigid grouping of children by birth year stops being a simplification and starts being a bottleneck. Progress aligned to readiness rather than age isn't a theoretical aspiration any longer. It's a demonstrated capability.
The implications for teachers are the opposite of what the anxiety narrative suggests. If AI manages the brute-force sequencing of tailored content delivery, the human educator is freed for work no algorithm replicates: mentorship, moral framing, the spark of genuine curiosity. Not the fading authority in front of a whiteboard, but a conscious guide within a system built for growth.
The students who benefited most from the personalised track were the ones who would have struggled most on the fixed one. The floor was raised. The ceiling was removed. It's not optimisation - it's a different contract between education and the child.
We've spent decades renovating a building whose foundations belong to another era. The question is no longer whether the technology exists to replace it. The question is what we're waiting for. It’s hard to justify anything less for my daughter now that we know this works.
Join the conversation on LinkedIn →
← All writing
