Use AI safely in the classroom: a teacher's guide

Start with a school-approved tool, a clear learning goal and work students can complete without AI. The steps below are a suggested teaching routine informed by research and UNESCO guidance, not a classroom trial we conducted.

The screenshots show public guidance and a source-checking example. They contain no student data.

1. Get approval before involving students

1. Ask your school's lead which tools and account types students may use. Check the provider's current age rules and any required parent permission.
2. Confirm what happens to prompts and uploads: who can access them, how long they are kept and whether they may train models.
3. Use invented examples. Keep names, faces, grades, health details and identifiable student work out of a general chatbot.
4. Prepare an equivalent activity for students who cannot use the tool.

UNESCO's guidance (opens in a new tab) recommends privacy protection, age-appropriate use and educational checks. It is guidance, not permission to override school policy or local law. If student use is not approved, use ordinary teaching materials. Staff preparation with AI also needs approval.

School photos deserve particular care. eSafety's school imagery advisory (opens in a new tab) explains the risk of image misuse and why schools should review sharing decisions.

Enlarge image

eSafety school photo checklist covering purpose, consent and identifying details

×

eSafety school imagery checklist. Text © Commonwealth of Australia, CC BY 4.0. Cropped screenshot.

Open full-size image

2. Ask for a hint after an initial attempt

1. Pick one goal, such as explaining a step in an equation.
2. Ask students to attempt the task first.
3. Test the exact task yourself before class. Check the facts, difficulty and feedback.
4. If the tool is approved, start with a prompt such as: "I have tried this problem. Ask me what I did, then give one hint. Wait for my next attempt before continuing."

In a randomized high-school maths study in Turkey, students using a basic GPT-4 tutor did better during assisted practice but worse on the subsequent exam without it. A tutor with learning safeguards avoided that observed penalty; it did not produce a statistically detectable improvement on the unassisted exam. Bastani et al., 2025 (opens in a new tab)

That protected tutor included teacher-written solutions and feedback guidance. Our example prompt does not reproduce the tested system or guarantee that a chatbot will withhold answers. Stop and intervene if it supplies the solution or gives misleading feedback.

3. Check one claim against a real source

1. Ask students to mark one factual claim in the reply.
2. Open a textbook or the relevant scientific institution's page directly. A link supplied by AI still needs checking.
3. Find the passage that supports the claim. Record the source and any correction.

For example, open NASA's What Is Earth? (opens in a new tab) and select Why Do We Have Day and Night? in Contents. Check whether the explanation connects day and night to Earth's spin. Read what it says about day and night near the poles. A plausible answer or a convincing citation alone is insufficient.

Enlarge image

NASA explanation of day and night with the polar daylight exception

×

NASA’s What Is Earth? explanation, including the polar exception. Cropped source-checking screenshot.

Open full-size image

4. Finish with a task without generative AI

1. Close the chatbot. Retain students' agreed accessibility accommodations.
2. Give a new question about the same concept.
3. Ask students to explain their reasoning. Use the result to decide what needs reteaching.

A separate randomized study of 194 Harvard undergraduates found higher immediate physics post-test scores with a carefully designed AI tutor than with the comparison active-learning lessons. It covered two lessons, used expert-written solutions and did not establish long-term benefits for schoolchildren. Kestin et al., 2025 (opens in a new tab)

The studies tested specific tutors and teaching methods. Their results cannot be assumed to apply to other chatbots or teaching settings. A polished AI-assisted submission alone cannot show what a student can do independently.

5. Keep assessment decisions with the teacher

State which uses are allowed for the assignment and how students should disclose them. Review their drafts, sources and explanation of the work before deciding whether a rule was broken.

Do not treat an AI-detector score as proof of cheating. A 2023 evaluation of seven detectors found false accusations in human-written essays by non-native English writers. That finding concerns the tools and samples studied, not a measured error rate for today's products. Liang et al., 2023 (opens in a new tab)

If a response is harmful or exposes private information

Stop the activity, support the student and follow the school's safeguarding or data-incident procedure. Keep access to any necessary record restricted. Never save or forward nude or sexual imagery of children, as eSafety's advisory explains. Check the tool before resuming.

Papers and guidance

• Bastani et al. (2025), Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS. Paper and DOI (opens in a new tab).
• Kestin et al. (2025), AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting. Scientific Reports. Paper and DOI (opens in a new tab).
• Liang et al. (2023), GPT detectors are biased against non-native English writers. Patterns. Published paper (opens in a new tab); open author version (opens in a new tab).

Official guidance: UNESCO, 2023 (opens in a new tab), eSafety's school imagery advisory (opens in a new tab).

Screenshot text: eSafety Commissioner, CC BY 4.0 (opens in a new tab). Screenshots are cropped; no endorsement is implied.