AI/Paper/25/Generative AI Without Guardrails Can Harm Learning

• Generative AI Without Guardrails Can Harm Learning: Evidence from High School Mathematics

  • PDF preprint
  • PNAS (2025) — cited in the 2025 Economic Report of the President

  • Key Findings

    • Turkish high school math students randomized to ChatGPT access for homework had better homework completion but worse performance on unassisted tests
    • Self-regulated AI use short-circuits the productive struggle required for learning

  • Authors

    • Hamsa Bastani, Osbert Bastani, Alp Sungu, Haosen Ge, Özge Kabakçı, Raphael Mariman ( equal first authorship) — Wharton / UPenn

  • Links

    • Cited in Person/Ethan Mollick/Blog/26/05/Choosing to Stay Human
    • Contrast with AI/Paper/25/Effective Personalized AI Tutors via LLM-Guided Reinforcement Learning — same team, adaptive tutoring improves outcomes