Here's what keeps me up at night about AI in education. It's not that students are cheating. It's that students can use AI, feel genuinely more confident, report that they learned more, be more satisfied with the experience, and still walk away knowing less. The feeling of mastery and actual mastery come apart. And AI widens that gap.

This isn't a hunch. Two studies published this year put hard numbers on it, and together they tell a single, clear story that should shape how we teach and how we let our students learn. Spoiler: it’s NOT banning it.

This week I'm synthesizing two 2026 studies, because neither is as powerful alone as they are together. Both ask the same underlying question, from different angles, and arrive at the same answer.

Researchers at Sungkyunkwan University ran an experiment with 88 university students, splitting the learning process into three stages: understanding concepts, solving problems, and reviewing results (Joo, Lee, & Lee, 2026). Their findings landed on a genuine paradox. Using AI during the problem-solving stage improved test performance, likely by cutting unnecessary cognitive load. But students with high AI dependency told a different story: they scored lower on objective tests, while reporting higher perceived learning and higher satisfaction. They accepted AI answers uncritically, skipped the hard work of making the knowledge their own, and felt great about it. The researchers named this the illusion of mastery. As Professor Changjun Lee put it, relying on AI uncritically can produce the paradoxical result of lowering actual achievement while raising the feeling of success.

The second study is much larger, a multisite, cluster-randomized field experiment with 1,176 first-year university students across four universities and three science domains (Ates, 2026). It compared four ways of giving students feedback on their scientific writing: peer feedback only, direct AI feedback, reflective AI feedback (where students evaluate their own work before seeing the AI's critique), and a hybrid of all three.

The results are crucial. Direct AI feedback produced the fastest immediate improvement to a draft. But when researchers tested students later, on their own, with no AI allowed, the picture flipped. The reflective and hybrid designs, the ones that forced students to do their own evaluative thinking first, produced significantly better durable learning and better AI-free performance. In the authors' framing, the educational value of AI depended less on AI access itself than on whether the feedback design preserved student agency, judgment, and ownership.

Put the two studies side by side, and the message is unmistakable. AI can improve the immediate product and the immediate feeling. Neither guarantees learning. What determines whether AI helps or hinders learning is whether the student is still doing their own cognitive work or handing it over.

This is cognitive surrender versus cognitive challenge, the exact framework we've built this newsletter around, now confirmed in two independent 2026 studies with a combined sample of over 1,200 students.

Watch for "AI-free transfer" to become a standard way of measuring whether AI-assisted learning actually stuck. It's a simple, powerful idea: can the student still do it when the tool is taken away? That's the question that separates real learning from borrowed performance, and it's the one every OT program should be asking about its own graduates.

This is the blueprint for how we train students & clinicians. Our students are building clinical reasoning, the capacity to evaluate, judge, and decide, that they will use, unassisted, in front of real patients for decades. If they learn by accepting AI outputs uncritically, they may pass the assignment, feel confident, and never build the reasoning the credential is supposed to certify. The illusion of mastery in a classroom becomes a competence gap at the bedside.

The second study is the hopeful part, and it's genuinely actionable. It tells us exactly how to integrate AI without hollowing out learning: make students do their own thinking first, then let AI serve as a comparison. Have a student write their own clinical reasoning, their own assessment interpretation, their own intervention rationale, before they consult AI. Then AI becomes a way to check and sharpen their judgment, not replace it. That's not a restriction on AI. It's the design that makes AI genuinely educational. And it's exactly what our AI in Practice section this week is built around.