How to Use an AI Tutor So It Helps: What the Studies Agree On

Two AI tutors were tested in the same Turkish high school. One raised practice scores 48% and then left students 17% worse on an exam taken without it. The other raised practice scores 127% and did no harm. Same AI model underneath. The difference was how it was set up. That single result, from a 2025 study in PNAS, is the best short answer to how to use an AI tutor: the design decides whether your child learns or just finishes. Read alongside a Harvard trial where a well-built AI tutor beat an active-learning class, and Stanford trials where kids barely used theirs, the research converges on a handful of rules you can apply at your kitchen table tonight.

Key Takeaways

• The guarded tutor in Bastani et al. (PNAS, 2025) gave hints, not answers, and was fed teacher-written solutions; it avoided the 17% exam drop seen with a plain chatbot.
• Harvard’s Kestin et al. (Scientific Reports, 2025) found a carefully structured AI physics tutor helped students “learn significantly more in less time” than an active-learning class.
• Stanford trials (2026) found elementary students used an AI reading tutor only 2 to 5 minutes a week; a human checking in raised engagement 71 to 80%.
• Students in AI groups consistently misjudged their own learning, so “it helped me” isn’t evidence. Test without the tool.
• Five rules follow: hints not answers, ground it in correct material, the child does the hard step, a human builds the routine, and check learning with the AI closed.

How to use an AI tutor: what the strongest studies tested

There are very few rigorous studies of AI tutoring. Stanford SCALE’s 2026 review of AI in K–12 found more than 800 relevant papers but only 20 with strong causal evidence. Its summary of what those show: AI improves performance while students have access; effects “when assessed independently without AI support… are mixed”; “easier doesn’t mean better”; and “tools designed with pedagogical guardrails… show more promise than general purpose AI tools.”

Three studies do most of the work in this article.

The Turkish math study. Hamsa Bastani and colleagues ran a field experiment with nearly 1,000 high school students, published in PNAS in 2025. “GPT Base” was a plain chat interface on GPT-4. “GPT Tutor” used the same interface with hidden instructions to give hints instead of answers, plus teacher-written solutions and common mistakes for each problem.

The Harvard physics study. Gregory Kestin and colleagues ran a randomized trial with 194 undergraduates, published in Scientific Reports in 2025. Their tutor was designed around seven research-based teaching practices, including active learning, managing cognitive load, promoting a growth mindset, scaffolding, accurate explanations, timely feedback and self-pacing.

The Stanford reading studies. Carly Robinson and colleagues at Stanford SCALE ran two randomized trials with 355 students in grades 1 to 5, reported in 2026 in a working paper not yet peer-reviewed. Some students used an AI reading platform alone; others used it with a human helper focused on engagement.

Study	Students	AI design	Main result
Bastani et al., PNAS 2025	~1,000 high schoolers, Turkey	Plain chatbot vs. hint-only tutor with teacher solutions	Plain: +48% practice, 17% lower exam. Hint-only: +127% practice, no exam harm
Kestin et al., Sci. Reports 2025	194 Harvard undergraduates	Custom tutor built on seven teaching practices	Learned significantly more in less time than active-learning class
Robinson et al., working paper 2026	355 students, grades 1–5	AI reading platform, alone vs. with human helper	Used 2–5 min/week; human helper raised stories completed 71–80%; no reading gains
Fan et al., BJET 2024	117 university students	ChatGPT vs. human expert vs. checklist vs. nothing	ChatGPT group improved essays most; knowledge gain no different

Rule 1: The tutor gives hints, not answers

This is the clearest finding in the literature. In the Turkish study, the most common message students sent the plain chatbot was “What is the answer?” The authors concluded students used it as a “crutch.” The hint-only tutor removed that option, and the exam harm largely disappeared.

The mechanism is simple. Learning requires retrieval and effort: the child has to produce the next step. An AI that produces it for them does the part that builds the skill.

You can approximate this at home. Start any tutoring chat with an instruction like: “You are a tutor. Never give me the final answer. Give one hint at a time, and ask me what I think the next step is.” It won’t be as reliable as a purpose-built tutor; chatbots don’t always follow instructions. But it shifts the default.

Rule 2: Ground it in correct material

General chatbots are often wrong on schoolwork. When the Bastani team asked the plain chatbot for answers to its 57 practice problems, ten times each, it was right only 51% of the time. The hint-only tutor rarely erred because its hidden instructions contained the correct solutions.

Kestin’s team found the same thing from the other direction. A system prompt alone “could not reliably provide enough structure,” so they “enriched our prompts with comprehensive, step-by-step answers.” Preparing those materials for a lesson “was done over a few days.”

At home: give the AI the source. Paste in the textbook section, the worked example from class, or the teacher’s study guide, and tell it to tutor from that. Prefer school-provided tools built on the curriculum over open chatbots for subjects where accuracy matters.

Rule 3: The child does the hard step, and someone checks afterward

The 2024 British Journal of Educational Technology study by Yizhou Fan and colleagues found that university students using ChatGPT for a writing task improved their essays more than any other group, but their “knowledge gain and transfer were not significantly different.” The authors warned of “metacognitive laziness”: skipping the planning, monitoring and self-checking that build skill.

There’s a second trap. In the Turkish study, students in both AI groups misjudged their learning. The plain-chatbot group did worse and didn’t notice; the hint-only group did no better on the exam but believed they did significantly better. So a child’s sense that “it really helped” is not reliable.

The fix is a check with the tool closed. After an AI session, ask for one similar problem done alone, or a one-minute explanation of the idea. That mirrors the studies’ unassisted exams.

Rule 4: A human builds the routine

The Stanford reading trials are the reality check on everything above. A well-designed tutor does nothing if a child doesn’t use it. Students scheduled for 60 minutes a week averaged 2 to 5. About half in the independent group never logged on.

A human helper who checked in, fixed tech problems and set norms raised the number of stories completed by 71 to 80%. It didn’t raise reading scores, likely because usage still fell far short of the recommended dose. Sal Khan described the same problem with Khanmigo: “Some will; most won’t” ask an AI for help on their own. We cover that in our article on the Stanford AI reading tutor findings.

For parents, this means short, scheduled sessions with you nearby at the start, rather than “use it whenever you’re stuck.”

Rule 5: Judge it by what your child can do without it

Every study above measured learning without the AI present, because that’s what counts. Use the same standard. Homework quality is not the metric. Quiz scores, in-class work, and your child’s ability to explain the material are. The OECD’s 2026 conclusion is the same: better homework is not proof of learning.

How to Teach Your Kid About Using an AI Tutor Well

Kids who understand why the rules exist follow them more willingly.

Ages 5–8: the “hint jar” game

Write three hint cards for a simple puzzle, such as a 20-piece jigsaw or a maze. Your child can take one hint at a time from a jar, but you never solve it for them. Afterward, ask: “Did the hints help you do it, or did they do it for you?” You’re teaching the difference between help and doing, which is the core of Rule 1. Young kids generally shouldn’t use chatbots unsupervised; this game builds the idea first.

Ages 9–12: set up the tutor together

Sit with your child and write the tutoring instruction together: “Don’t give me the answer. Give one hint. Ask what I think.” Paste in the worksheet example from class. Let them do three problems this way. Then close the laptop and have them try one more alone. Talk about how it felt: harder than asking for answers, and why that’s the point.

Ages 13+: run a two-tutor experiment

Have your teen pick a topic they’re studying and two short problem sets of similar difficulty. On the first set, they use a chatbot however they like. On the second, they use it in hint-only mode with the class notes pasted in. Two days later, they do three new problems with no AI and see which method left them better prepared. They’ve replicated, in miniature, the design of the PNAS study.

The question to ask: “If you can’t do it with the AI closed, who actually learned it?”

Setting up an AI tutor at home: the practical checklist

Use a tutor mode when one exists

Many school platforms and some chatbots offer a “study” or “learning” mode that withholds answers. Prefer it. If none exists, use the hint-only instruction above.

Feed it the right source

Paste in the textbook passage or class example. Ask it to tutor from that material and say when something isn’t covered.

Schedule it, short and regular

Two or three 15-minute sessions a week, at a set time, beat open-ended access. Sit nearby for the first few minutes.

Close the laptop and check

End every session with one problem or explanation done without the AI. Note how it goes over a few weeks.

What not to do: hand over a general chatbot and walk away

That is the exact condition with the clearest evidence of harm: unguarded answers on demand, judged by homework quality, with no human in the loop. Our guide to AI tutors for homework help, ages 8 to 14 has more on choosing tools.

What to Watch For Over the Next 3 Months

• Week 4: After AI sessions, can your child do one problem alone? Keep a simple note: yes, partly, no.
• Month 2 red flags: Your child asking for the answer as soon as a problem gets hard, switching from tutor mode to a chatbot that gives answers, or homework improving while quizzes don’t.
• Month 3 self-check: Do test scores without AI match what you see in AI-assisted homework? Is your child using AI for explanations and hints rather than answers? Is there a regular routine with a human checking in?

Frequently Asked Questions

How do I make an AI tutor actually help my child learn?

Set it to give hints instead of answers, give it correct source material such as class notes, make sure your child does the hard step, keep sessions short and scheduled with you checking in, and end each session with a problem done without the AI. These rules come from the strongest randomized studies of AI tutoring.

Do AI tutors work?

Sometimes. A hint-only tutor grounded in teacher solutions avoided learning harm in a 1,000-student PNAS study, and a carefully designed Harvard physics tutor outperformed an active-learning class. But a plain chatbot lowered later exam scores 17%, and elementary students in Stanford trials barely used their AI reading tutor. Design and use decide the outcome.

Should my child use ChatGPT as a tutor?

It can work for older kids if set up carefully: instruct it to give hints only, paste in class material so it doesn’t invent answers, and check learning afterward without it. General chatbots were right only about half the time on the Turkish math problems, so accuracy is a real concern. For younger children, supervised use or school-provided tools are safer.

My child says the AI tutor helps a lot. Is that enough?

No. In the PNAS study, students in both AI groups misjudged their learning; one group did worse and didn’t notice, the other believed it did better when it hadn’t. The reliable test is performance without the AI: one similar problem or a quick explanation after each session.

How much time should a child spend with an AI tutor?

Studies don’t give an exact number for every age. Short, regular sessions with clear goals work better than unlimited access. In Stanford’s reading trials, recommended use was about 30 minutes a week, and most children fell far short of it without a person building the routine.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. 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). https://doi.org/10.1073/pnas.2422633122
2. Kestin, G., Miller, K., Klales, A., Milbourne, T., & Ponti, G. (2025). “AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting.” Scientific Reports, 15. https://doi.org/10.1038/s41598-025-97652-6
3. Robinson, C., et al. (2026). “Access Is Not Enough: Human Support Improves Engagement with AI Tutoring.” Working paper, Stanford SCALE Initiative / National Student Support Accelerator. https://nssa.stanford.edu/studies/access-not-enough-human-support-improves-engagement-ai-tutoring
4. Fan, Y., et al. (2024). “Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance.” British Journal of Educational Technology. https://doi.org/10.1111/bjet.13544
5. Fesler, L., Martinez Claeys, J. P., Agnew, C., & Loeb, S. (2026). “The Evidence Base on AI in K-12: A 2026 Review.” Stanford SCALE Initiative. https://scale.stanford.edu/sites/default/files/The%20Evidence%20Base%20on%20AI%20in%20K-12%20Report.pdf
6. Barnum, M. (2026, April 9). “Why Sal Khan’s AI revolution hasn’t happened yet, according to Sal Khan.” Chalkbeat. https://www.chalkbeat.org/2026/04/09/sal-khan-reflects-on-ai-in-schools-and-khanmigo/