Evidence review · Updated October 2026

What actually helps kids learn with AI

Some studies say AI tutoring doubles learning. Others say it hurts. Both are true, and the difference is design. Here is what the research shows, how strong each finding is, and what it means for a classroom.

For teachers and leaders8 minute readStrength rated on every finding

The short version

The tool matters less than the task design.

  • AI that hands over answers raises practice scores and can lower later, unaided learning.
  • AI that makes students try first, gives hints instead of answers and is checked by a teacher looks much safer.
  • Much of the alarming brain research is on adults, with small samples. Treat it as a warning, not a verdict.

How to read this page

Not all evidence is equal

Each finding carries a label so you can see how much weight it can bear.

Stronger

Randomized or controlled study with students, published in a peer-reviewed venue.

Suggestive

Controlled but small, short-term, vendor-linked, adult-only, or not yet peer reviewed.

Early or correlational

Surveys, brain-imaging pilots or preprints. Raises questions but does not settle them.

Research method note: this review was compiled in October 2026 from published abstracts, summaries and reports. Some primary papers could not be opened directly, so we cite headline findings only and link every source. Please open the sources before quoting a number.

The studies

What was found

Bastani et al. · PNAS 2025

Stronger one school, one subject

High-school math students practiced with (a) no AI, (b) a ChatGPT-style assistant or (c) a tutor version that gave hints and used teacher-written solutions. During practice, scores were about 48% higher with the plain assistant and 127% higher with the tutor. On a later exam without AI, the plain-assistant group scored about 17% lower than the control group. The tutor group’s drop was largely eliminated.

Keep in mind: A single school in Turkey over a few sessions. PNAS published a correction that fixed an author affiliation; the reported results were unchanged. Quote figures from the current version.

pnas.org/doi/10.1073/pnas.2422633122

Kestin et al. · Scientific Reports 2025

Suggestive

In a Harvard physics course (194 students), a carefully engineered AI tutor outperformed in-class active learning: median learning gains were more than double, and students spent less time. The tutor had expert-written solutions built into its instructions and guided one step at a time.

Keep in mind: University students, two physics topics, short-term. Not evidence about children.

nature.com/articles/s41598-025-97652-6

Tutor CoPilot · Stanford

Suggestive

AI that coaches the human tutor (not the child) was tested with 900 tutors and about 1,800 K–12 students. Students were about 4 percentage points more likely to master a topic, and about 9 points for lower-rated tutors, at roughly $20 per tutor per year.

Keep in mind: Measured on lesson-end mastery checks in an online tutoring program, not in whole-class instruction.

arxiv.org/abs/2410.03017

Liu et al. · EdWorkingPapers 2026

Early · working paper

In a randomized university trial with 2,379 undergraduates, access to a course-integrated AI tutor reduced final grades (about 0.37 standard deviations) and course participation. The losses were larger for first-generation students.

Keep in mind: We could only confirm the headline results. Tutor design details were not verified. It shows that “AI tutor” alone does not guarantee help.

edworkingpapers.com/ai26-1598

Kosmyna et al. · MIT Media Lab 2025 (“Your Brain on ChatGPT”)

Early · preprint

54 adults wrote essays alone, with a search engine, or with ChatGPT while wearing EEG caps. Brain-connectivity measures were lower with more tool support, and the ChatGPT group had more trouble quoting their own essays. The authors call the pattern “cognitive debt.”

Keep in mind: Adults, not children. Small sample. Preprint when released. The name “cognitive debt” is an interpretation, and a 2026 commentary notes adolescents were not studied.

arxiv.org/abs/2506.08872

Lee et al. · Microsoft Research and CMU, CHI 2025

Early · survey

In a survey of 319 knowledge workers, people who trusted AI more reported using less critical thinking, while people with more confidence in themselves reported more. Their effort moved from doing the task to checking and supervising it.

Keep in mind: Self-reported by adult workers. Correlation, not cause.

Microsoft Research PDF

The pattern

What separates AI that helps from AI that harms

The strongest studies point to the same design choices.

1

Attempt first

Students write or solve before the tool speaks. This uses the generation effect: you remember what you produce.

2

Hints, not answers

Guardrailed tutors that withheld final answers avoided most of the learning loss in the Bastani study.

3

Verified content inside

Teacher-checked solutions built into the tutor prompt reduce wrong hints (used in both Bastani and Kestin).

4

A human in the loop

Tutor CoPilot helped the adult help the child. The teacher stays the teacher.

5

Unaided check afterward

Measure learning without the tool. Practice scores with AI can look great while learning is flat or falling.

6

Stay humble

The 2026 Liu result shows that a well-meant tutor can still hurt. Pilot, measure and adjust.

Same lesson, three ways. A student asks for help with a math problem. Answer machine: “The answer is 408.” Guardrailed tutor: “What is 17 × 20 first? Try it and tell me.” Teacher-guided: students try on paper, the AI offers one hint, and the teacher checks understanding with a no-AI question. The last two keep the thinking with the student.

Learning science

Old findings that still win

These are well-established ideas from cognitive psychology. They apply to AI lessons just as they do to any other.

PrincipleEvidence anchorIn an AI lesson
Retrieval practiceDunlosky et al. 2013: practice testing rated high utilityStart by having students write from memory how the model works. Then check.
SpacingDunlosky et al. 2013: distributed practice rated high utilityRevisit last week’s idea in a three-minute warm-up, not only once.
Generation effectSlamecka & Graf 1978Students predict or write an answer before seeing the AI’s.
Productive failureSinha & Kapur 2021: 53 studies, small to moderate average benefit for struggle-then-instruction (g = 0.36), especially with careful design. In grades 2–5 and on domain-general skills, instruction-first came out ahead.For older students and subject-specific problems: ten minutes of wrestling with a problem before any explanation, by a person or an AI. Adapt the order to the learner, their prior knowledge and the goal.
MetacognitionBastani found students overestimated learning with unguarded AIAsk “How sure are you, and how will you check?” before and after using AI.
Self-explanationDunlosky et al. 2013: moderate utility“Explain, in your own words, why the AI’s answer is right or wrong.”

Bloom’s “2 sigma” result for one-to-one tutoring is often quoted for AI tutors. Treat it as an upper bound, not a promise: later reviews of human tutoring found smaller effects.

Checking what AI says

Lateral reading works

Professional fact-checkers do not stay on a page to judge it. They leave and open new tabs to see what others say about the source. This is called lateral reading, and it is the core of the SIFT method (Stop, Investigate the source, Find better coverage, Trace to the original).

In a randomized study in an urban school district (Wineburg et al., Journal of Educational Psychology 2022), students who took six lateral-reading lessons improved significantly more at judging online credibility than a comparison group.

No study we found tested lateral reading on chatbot output specifically. Applying it to AI answers is a reasonable inference, which is why our Verify It lab treats every AI answer as a lead to check, not a fact to accept.

Try it

Verify It lab

Four interactive scenarios per grade band, with feedback that explains why.

Start the lab →

Put it to work

Seven moves for Monday morning

  1. Paper first. Every AI task starts with a student’s own attempt, even a rough one.
  2. Name the AI level. Tell students whether the task is Level 0 (no AI) through Level 4. See the levels.
  3. Use hint-style tutors. Paste a tutor prompt from our toolkit that refuses to give final answers.
  4. Add an unaided checkpoint to every unit.
  5. Collect an AI receipt with any AI-assisted work.
  6. Grade the process: the plan, the checking and the revision.
  7. Pilot small. Try one class, measure learning without AI, then decide.