When the answer arrives too soon.
A preregistered randomized trial in a Turkish high school compared a standard GPT-4 chat interface, a teacher-informed GPT tutor with guardrails, and no generative-AI resource during mathematics study sessions.
- Maturity
- Demonstrated, stage 3 of 4
- Support
- 1 source · paper
- Evidence detail
- How we know ↓
A tool that makes practice easier can also let a learner bypass the thinking needed to solve a problem alone.
Widely available chat interfaces are designed to be helpful in the moment, not necessarily to protect learning.
The study compared GPT-4 assistance that resembled ordinary chat with a version whose prompts and boundaries were designed with teacher input to support learning.
What was shown, and what wasn’t
Shown
The paper's preregistered primary analysis of unassisted exams found that generative AI without guardrails could harm learning. The teacher-informed guardrailed tutor largely mitigated that negative effect.
Not shown · limits
- The study examined short-term exam outcomes in one high-school mathematics setting.
- The paper says the guardrailed tutor mitigated harm; it does not establish a complete or universally effective learning design.
- Long-term retention and transfer were not measured.
Bright editorial interpretation
It makes the design of help visible: a system can improve assisted performance yet leave less learning when the assistance disappears.
Still open
Which guardrails work across subjects and ages?
Can a tutor proactively find misconceptions without taking over the work?
How should educators assess learning when students have broad access to general chat tools?
How we know1 source · checked 2026-09-07 · no corrections
Original sources
Institutions: University of Pennsylvania · Budapest British International School
- Maturity
- Demonstrated
- Source published
- 2025-06-25
- Captured
- 2026-09-07
- Last source review
- 2026-09-07
- Editorial method
- AI-assisted source review
Bright compared this account with the linked original and supporting sources and kept reported, budgeted, projected, and observed claims distinct. Bright did not independently audit the underlying records.
Maturity describes the tested or operational setting. Confidence describes support for the particular claim; one does not determine the other.
Revision & correction history
2026-09-07 · It makes the design of help visible: a system can improve assisted performance yet leave less learning when the assistance disappears.
No corrections recorded.