In 2025, Wharton researchers published in PNAS an experiment with about a thousand high school math students. One group studied with access to a generic language model; another, with a version that carried pedagogical safeguards — it gave hints, asked for the student's attempt, did not hand over the ready answer; a third studied without AI.
The result
With the AI on, the first group performed 48% better on the exercises. On the test without AI, it performed 17% worse than those who never used it. The group with safeguards performed 127% better with the tool on — and, without it, had no drop relative to those who never used it.
The reading is simple and uncomfortable: an AI that does it for you produces performance while it is present and leaves less than it found when it goes away. An AI that teaches you to do it does not compromise what stays.
Beatriz's condition
Beatriz was built on the second condition. She does not hand over the answer: she builds the whiteboard, asks questions back, asks for your attempt and ends every session with a task for your real work. The weekly quota does not roll over — when it's used up, you practice and come back. Every three weeks, a checkpoint. At the end, an autonomy test taken with her out of the picture and a 30-to-50-day interval.
- Teaching you to do it, not doing it for you.
- Practice outside the platform as part of the method.
- A test with the AI removed, measured by the LUMEN standard.
The number that matters is not how much you perform with Beatriz at your side. It is how much stays when she leaves.
- Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., Mariman, R. “Generative AI without guardrails can harm learning”. PNAS 122 (2025). Figures checked against the primary source on July 22, 2026.
Feel free to quote — with the source.