Hamsa Bastani, Osbert Bastani, Alp Sungu, Haosen Ge, รzge Kurucu, Rehema Mushi (2025) โ Proceedings of the National Academy of Sciences (PNAS), 122(26). doi:10.1073/pnas.2422633122.
๐ Full text (PNAS) โ RCT preregistered with the University of Pennsylvania IRB; anonymized data and code at github.com/obastani/GenAICanHarmLearning
Summary
This landmark field experiment is among the first randomized controlled trials to causally demonstrate that unguarded generative-AI tutoring can harm skill acquisition, not merely fail to help. Conducted with nearly 1,000 high-school math students across ~50 classes at a large school in Turkey (Fall 2023โ2024), the study compares three arms assigned at the classroom level: a control arm (textbook/notes only), GPT Base (a ChatGPT-like GPT-4 chat interface), and GPT Tutor (GPT-4 with teacher-designed guardrails โ hints instead of answers, plus the correct solution and common mistakes baked into the prompt). Each 90-minute session had three parts: a teacher lecture, an AI-assisted practice period (treatment only here), and an unassisted closed-book exam on conceptually similar problems.
The core finding: access to generative AI sharply improves performance during practice but, without guardrails, degrades learning once the tool is removed โ a tradeoff driven by students using the AI as a "crutch" to copy answers rather than to learn.
Key Findings
Performance vs. Learning Tradeoff (intention-to-treat)
Relative to control (practice mean 0.284; exam mean 0.321, normalized grades):| Arm | Assisted practice (perf) | Unassisted exam (learning) |
|---|---|---|
| GPT Base | +48% (ฮฒ = 0.137\\) | โ17% (ฮฒ = โ0.054\*) |
| GPT Tutor | +127% (ฮฒ = 0.361\\) | โ 0 (ฮฒ = โ0.004, n.s.) |
- GPT Base significantly improves practice performance but significantly worsens later unassisted exam performance โ students who never had AI access actually outperformed them.
- GPT Tutor's guardrails largely eliminate the learning harm (point estimate near zero), though they do not produce a positive learning effect either.^[raw/papers/pnas-2025-guardrails-harm-learning.md]
Mechanism: Students Use GPT Base as a Crutch
Two candidate explanations for the harm were tested โ (1) GPT Base's errors mislead students, and (2) students offload thinking by copying answers. The evidence strongly favors the second:- Error analysis: When GPT Base's logical-error rate on a practice problem is higher, it hurts practice performance but shows no spillover to the corresponding unassisted exam problem โ so students aren't being systematically misled into exam errors.
- Engagement analysis: Students in the GPT Base arm send far fewer messages and overwhelmingly just "ask for the answer" or restate the question (superficial conversations dominate). In the GPT Tutor arm, a growing share of conversations are substantive (asking for help, attempting answers independently), and this improves within the very first session.
- GPT Base answered correctly only 51% of the time on the 57 practice problems (42% logical errors, 8% arithmetic errors) โ yet students still copy its outputs.^[raw/papers/pnas-2025-guardrails-harm-learning.md]
Students Don't Perceive the Harm
Students in the GPT Base arm performed worse on the exam but did not report learning or performing less; GPT Tutor users perceived they performed better than control even though exam scores were statistically indistinguishable. This perceived-vs-actual-learning mismatch parallels the "feeling of learning" literature and means self-report is an unreliable gauge of AI's learning impact.^[raw/papers/pnas-2025-guardrails-harm-learning.md]Secondary Results
- Skill-gap narrowing is temporary: Both AI arms reduced grade dispersion (HHI) during practice (biggest help to weakest students), but the effect does not persist on the unassisted exam.
- Limited heterogeneity: Little evidence of differential effects by student ability, resources, or effort on exam performance.
- Robustness: Intention-to-treat (including noncompliers), alternative specifications, and absenteeism checks all confirm the pattern.^[raw/papers/pnas-2025-guardrails-harm-learning.md]
Design Lesson: What the Guardrails Did
GPT Tutor differed from GPT Base in two ways: (1) the prompt instructed it to give hints, not answers, and (2) it was seeded with teacher-authored problem-specific information (correct solution, common mistakes, feedback guidance) โ making its hints accurate and checkable. This labor-intensive prompt design is what neutralized the crutch effect. The authors note GPT Tutor remains passive (it doesn't proactively probe misconceptions) and call for combining pedagogical software tutors with generative AI, plus "co-pilot" models that assist human tutors rather than replace them.^[raw/papers/pnas-2025-guardrails-harm-learning.md]Implications
- For the wiki: This is the strongest causal, field-deployed evidence for the over-reliance and cognitive-offloading phenomena, and a direct, preregistered demonstration of the genai-performance-vs-learning distinction (performance โ, learning โ). It complements generative-ai-reduced-study-time-math (population-scale behavioral evidence) and ai-assistance-reduces-persistence (lab causal evidence), and motivates scaffolding and socratic-ai-dialogue designs that preserve learning.
- For practice/policy: Because ChatGPT-like tools are already widely used by students off-platform, the GPT Base harm is immediately relevant. Deployments where learning matters need explicit guardrails (hint-not-answer prompts, teacher-authored scaffolding, answer-checking).
Limitations (per authors)
Single topic (math), single high school in Turkey, Fall 2023 (early GPT-4 era), short-term outcomes only; writing and other subjects lack the objective grading used here. Generalizability to newer models and other contexts requires further study.^[raw/papers/pnas-2025-guardrails-harm-learning.md]Related Pages
- genai-performance-vs-learning โ The performance-vs-learning distinction this RCT causally demonstrates
- over-reliance โ The "crutch" mechanism quantified here in a real classroom
- cognitive-offloading โ Closely related; offloading thinking to the AI
- generative-ai-reduced-study-time-math โ Population-scale (3.2M interactions) companion evidence of learning loss
- ai-assistance-reduces-persistence โ Lab causal evidence that AI assistance impairs unassisted performance
- scaffolding โ Guardrail design that preserves learning
- socratic-ai-dialogue โ Hint-not-answer tutoring pattern embodied by GPT Tutor
- intelligent-tutoring โ Where guardrailed generative tutors fit
- k-12 / stem-education โ Deployment context (high-school math)
- RCT โ The study design used
- learning-gains โ The outcome harmed by GPT Base
- student-experience โ Students' misperception of their own learning