Can an AI tutor help my child learn, or will it just do the work?
Short answer: an AI tutor can help a child learn, but only when it is set up to make them think. In the largest field study so far, students who practised math with a plain chatbot did better during practice and then scored 17% worse on the exam than students who never had it. Students using a version told to give hints instead of answers did not lose that ground.
What happened when students got ChatGPT for math practice?
In 2023, researchers ran a field experiment with nearly a thousand high-school students in Turkey (Bastani et al., PNAS, 2025). Some students practised with a standard GPT-4 chat window ("GPT Base"). Others got "GPT Tutor", the same model with instructions to give hints, check the student's steps, and withhold the final answer. A control group practised with textbooks only.
During practice, both AI groups solved far more problems than the control group. Then the AI was taken away for the exam. The GPT Base group scored about 17% lower than the control group. The researchers read the chat logs and saw why: many students asked for the answer, copied it, and moved on. The tool had become a crutch. GPT Tutor students largely avoided the drop, because the tutor made them do the steps.
The lesson for parents is simple: better homework is not the same as more learning. A tool that removes all difficulty can feel helpful while quietly removing the part that builds memory.
Is there evidence AI tutoring can work well?
Yes, when the design is deliberate. Three studies are worth knowing:
| Study | Who | What they tested | Result |
|---|---|---|---|
| Kestin et al., Scientific Reports, 2025 | 194 Harvard physics students | An AI tutor built on the same teaching principles as the class, versus an active-learning lecture | Students learned significantly more in less time, and reported more engagement |
| Henkel et al., 2024 | About 1,000 students in grades 3–9, Ghana | Rori, a WhatsApp math tutor that gives a hint before a worked solution, two 30-minute sessions a week | Effect size of 0.37 on math growth after 8 months |
| Wang et al., Tutor CoPilot, 2024 | About 900 human tutors and 1,800 K–12 students | AI suggestions given to human tutors in real time | Students were 4 percentage points more likely to master topics; 9 points with less experienced tutors |
The common thread is that the AI asked questions, gave hints in steps and kept the student active. In Tutor CoPilot, the transcripts show tutors with the tool gave away fewer answers and asked more guiding questions.
Why does struggling a little help children remember?
Memory researchers call this desirable difficulty. Pulling an answer out of your own head strengthens it far more than reading it again. (We cover the evidence in why quizzes beat rereading.) A chatbot that supplies the answer replaces retrieval with reading. That is the mechanism behind the 17% drop.
It also explains a forty-year-old benchmark. In 1984, Benjamin Bloom reported that one-to-one tutoring lifted students about two standard deviations above classroom teaching ("The 2 Sigma Problem"). Later reviews found smaller but still large effects. In VanLehn's 2011 review, human tutoring came in around 0.79 standard deviations and the best step-by-step computer tutors around 0.76. Good tutors rarely lecture. They prompt, wait, and correct one step at a time.
How do I set up an AI tutor so my child still does the thinking?
- Write the rules into the first message. For example: "You are a patient tutor for a 9-year-old. Never give the final answer. Ask one question at a time. If I am stuck, give a small hint, then a bigger one. After I answer, ask me to explain why."
- Have them try first. Make it a house rule that your child writes their own attempt before opening the chat. The AI checks work; it does not start it.
- End with a no-AI check. Finish every session with two or three similar problems done without help. If those go badly, the session was copying, not learning.
- Sit nearby for the first few weeks. You will quickly see whether the chat is a conversation or an answer machine.
- Mind the age rules. Most general chatbots require users to be 13 or older, and teens need a parent's permission. For younger children, a parent drives and the child talks.
What can parents use AI for right now?
The safest high-value use is preparation rather than live tutoring. Ask an assistant to:
- write ten practice questions on tonight's topic, at three levels of difficulty;
- turn a textbook page into five flashcards with short answers;
- explain a concept to you in plain words, so you can explain it to your child;
- suggest a real-world example ("fractions while baking", "angles on a skateboard ramp").
Check what comes back. Language models make confident mistakes, especially with arithmetic and dates. Ethan and Lilach Mollick's guide Assigning AI: Seven Approaches for Students, with Prompts has good templates for tutor, coach and "teach the AI" roles.
The bottom line
AI does not make children learn more or less by itself. Set up to hint, question and check, it can be a strong practice partner. Set up to answer, it can make homework look better while learning gets worse. Keep the rule simple: the child does the thinking, the AI asks the questions.
DadSchool is built this way. Lessons are followed by flashcards and a test the child takes alone, and rewards come only from passing that test.
References
- Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS, 122(26). doi:10.1073/pnas.2422633122
- Kestin, G., Miller, K., Klales, A., Milbourne, T., & Ponti, G. (2025). AI tutoring outperforms in-class active learning. Scientific Reports. doi:10.1038/s41598-025-97652-6
- Henkel, O., Horne-Robinson, H., Kozhakhmetova, N., & Lee, A. (2024). Effective and scalable math support: Evidence on the impact of an AI-tutor on math achievement in Ghana. arXiv:2402.09809
- Wang, R. E., Ribeiro, A. T., Robinson, C. D., Loeb, S., & Demszky, D. (2024). Tutor CoPilot: A human-AI approach for scaling real-time expertise. arXiv:2410.03017
- Bloom, B. S. (1984). The 2 sigma problem. Educational Researcher, 13(6), 4–16.
- VanLehn, K. (2011). The relative effectiveness of human tutoring, intelligent tutoring systems, and other tutoring systems. Educational Psychologist, 46(4), 197–221.