CITATION — REFERENCE ENTRY

Generative AI without guardrails can harm learning: Evidence from high school mathematics — Proceedings of the National Academy of Sciences

Revision beff0b18-6997-483c-8724-998568230c99 · 9/26/2026, 3:06:46 PM UTC
Key
bastani-2025-generative-ai-guardrails
Authors
Bastani, Hamsa; Bastani, Osbert; Sungu, Alp; Ge, Haosen; Kabakcı, Özge; Mariman, Rei
Issued
2025-6-25
Type
article-journal
Container
Proceedings of the National Academy of Sciences
Volume
122
Issue
26
Pages
e2422633122
Raw CSL JSON
{
  "DOI": "10.1073/pnas.2422633122",
  "URL": "https://doi.org/10.1073/pnas.2422633122",
  "page": "e2422633122",
  "type": "article-journal",
  "issue": "26",
  "title": "Generative AI without guardrails can harm learning: Evidence from high school mathematics",
  "author": [
    {
      "given": "Hamsa",
      "family": "Bastani"
    },
    {
      "given": "Osbert",
      "family": "Bastani"
    },
    {
      "given": "Alp",
      "family": "Sungu"
    },
    {
      "given": "Haosen",
      "family": "Ge"
    },
    {
      "given": "Özge",
      "family": "Kabakcı"
    },
    {
      "given": "Rei",
      "family": "Mariman"
    }
  ],
  "issued": {
    "date-parts": [
      [
        2025,
        6,
        25
      ]
    ]
  },
  "volume": "122",
  "container-title": "Proceedings of the National Academy of Sciences"
}

Claims

  1. Students with unguarded GPT-4 access used it as a crutch during practice, often asking directly for answers, and then performed worse on their own.
    "Without guardrails, students attempt to use GPT-4 as a “crutch” during practice problem sessions, and subsequently perform worse on their own."
    Locator: section: Abstract · Quote language: en
  2. In a randomized trial with nearly 1,000 Turkish high school students, access to an unrestricted GPT-4 interface improved practice scores by 48% but reduced scores on a later unassisted exam by 17%; a tutor version with guardrails improved practice scores by 127% and largely eliminated the harm, without producing a positive effect.
    "students in the GPT Base arm perform statistically significantly worse than students in the control arm by 17%; this negative effect is essentially eradicated in the GPT Tutor arm, though we still do not observe a positive effect."
    Locator: section: Introduction · Quote language: en
  3. Students who copied solutions from GPT-4 did not perceive any reduction in their learning or performance.
    "students do not perceive any reduction in their learning or subsequent performance as a consequence of copying solutions"
    Locator: section: Introduction · Quote language: en
Available in