# 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.

Canonical: https://brightaifuture.com/discoveries/guarded-learning
Format: discovery
Source publication: 2025-06-25
Bright publication: 2026-09-07
Substantive update: None recorded
Evidence and review: Demonstrated; confidence: unassessed; approved; ai-assisted. AI-assisted editorial comparison with the cited primary source; result, setting, source date and limitations retained. Independently checked within the research team. Publication authorized by the site owner; no human source review is claimed.

## The human problem

A tool that makes practice easier can also let a learner bypass the thinking needed to solve a problem alone.

## The prior constraint

Widely available chat interfaces are designed to be helpful in the moment, not necessarily to protect learning.

## AI’s actual role

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.

## The documented result

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.

## Why it may matter

It makes the design of help visible: a system can improve assisted performance yet leave less learning when the assistance disappears.

## Limitations

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.

## Unresolved questions

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?

## Provenance and history

{
  "dates": {
    "eventDate": null,
    "publicationDate": "2025-06-25",
    "captureDate": "2026-09-07",
    "lastReviewedDate": "2026-09-07"
  },
  "provenance": {
    "origin": "editorial",
    "externalId": "https://doi.org/10.1073/pnas.2422633122"
  },
  "revisions": [
    {
      "id": "revision:a3f56235a1d3df554891",
      "recordedAt": "2026-09-07",
      "summary": "It makes the design of help visible: a system can improve assisted performance yet leave less learning when the assistance disappears.",
      "sourceIds": [
        "source-guarded-learning"
      ]
    }
  ],
  "corrections": []
}

## Original sources

- [Generative AI without guardrails can harm learning: Evidence from high school mathematics](https://doi.org/10.1073/pnas.2422633122)

## Continue exploring

- [Work & learning](https://brightaifuture.com/worlds/work)
- [How can AI expand learning while preserving human judgment?](https://brightaifuture.com/threads/learning)
- [When does help become a shortcut?](https://brightaifuture.com/experiments/learning-with-help)