Why This Study Matters
This is one of the first rigorous field experiments demonstrating that generative AI access can actively harm learning outcomes. The finding that students perform worse after losing AI access than students who never had it suggests a dependency effect with serious implications for education policy.
Research Design
Nearly 1,000 high school students were randomly assigned to three conditions:
- Control: No GPT access
- GPT Base: Standard ChatGPT-style interface
- GPT Tutor: GPT with pedagogical safeguards (e.g., Socratic prompting, hints instead of answers)
Students used the tools for a portion of their math curriculum, then were tested without access.
Key Insight
The critical finding is the asymmetry: AI boosts performance during use but degrades it after removal. This suggests students outsource cognitive effort to the AI rather than building understanding. However, the GPT Tutor condition — which guided students rather than giving direct answers — largely prevented this harm, pointing toward design principles for educational AI tools.
Bottom Line
Giving students raw access to generative AI without safeguards can create a cognitive crutch that leaves them worse off than if they had never used it. But thoughtfully designed AI tutoring (Socratic-style, with guardrails) can capture the benefits while protecting learning.