JOURNAL ARTICLE · EDITORIAL BRIEF
Generative AI without guardrails can harm learning: Evidence from high school mathematics
Hamsa Bastani; Osbert Bastani; Alp Sungu; Haosen Ge; Özge Kabakcı; Rei Mariman
PNAS · 25 June 2025
The finding or development
AI support improved practice performance, but unguarded assistance could undermine later performance without AI. A tutor using teacher-informed safeguards mitigated the harm.
What supports it
Peer-reviewed randomised field experiment comparing GPT-4-based support with and without learning safeguards and a no-AI condition in high-school mathematics.
Reviewed: Publisher abstract, significance statement and study-design excerpt; full methods not appraised.
Why it matters for design
Assess what you can explain independently after using assistance, as well as what you can produce with it.
Read with this in mind
School mathematics is not UX research. Transfer to qualitative interpretation or professional practice has not been established here. This brief is based on selected publisher sections, not a full methodological review.
Original source
Opens the publisher outside this reading room. If nothing opens, copy the source address below.
Reference tools · Paperpile & more
Using Paperpile? Open the original source above, then use the Paperpile extension to save it to your library. Its inline buttons appear on supported websites.
Paperpile’s browser extension guide (opens an external website)Or download the original reference and import the file into Paperpile, Zotero or EndNote. Check the details against the source before citing.
Citation details are requested from Crossref when you download. The file contains a reference, not the full paper.
DOI: 10.1073/pnas.2422633122
This is the complete UX Observatory brief, not the publisher’s full article. Follow the original source to read the work.