Do AI tutors actually help kids learn?

Sometimes, and usually by a small amount. Well-designed tutors have shown real gains in a few trials (Harvard physics, Nigeria, Ghana), but the biggest independent school trial, of Khanmigo, found small gains about the same as ordinary practice software. Stanford's 2026 review found only 20 strong studies and none with US K-12 students.

Can using ChatGPT for homework make my child learn less?

There is good evidence it can. In a trial with about 1,000 high schoolers, plain ChatGPT raised practice scores but students did 17% worse on a later test without it. Large platform data suggests the same pattern on easy-to-paste math problems. Guardrails that give hints instead of answers reduced most of the harm.

What makes an AI tutor good rather than harmful?

The studies that went well share a pattern: the AI gives one step or hint at a time, holds back the answer, is fed correct solutions, and steps in after a mistake. Tools built that way look better than general chatbots, though even good design does not help if kids do not use it.

Is an AI tutor as good as a human tutor?

Not yet as a replacement. Human tutoring several times a week has strong evidence (about a third of a standard deviation on average). AI-only tutoring has much thinner evidence, and Stanford's tutoring center says it does not yet meet that bar. The best-supported use right now is AI helping human tutors.

What is Bloom's "2 sigma problem" and is it true?

In 1984 Benjamin Bloom reported that tutored students did better than 98% of classroom students, a two standard deviation gain. Later reviews found real tutoring effects are much smaller, around 0.3 to 0.8, so promises that AI will deliver two sigma are not supported.

Are AI tutors new, or has this been tried before?

Computer tutors have been studied for over 40 years. Older step-by-step tutors like Cognitive Tutor, ASSISTments and AutoTutor had solid, moderate effects in many trials. Chatbots are new in how they talk, but the lessons about feedback, practice and usage carry over.

Why don't kids use AI tutors much, even good ones?

Usage is the biggest problem found so far. In the Khanmigo trial most students tried the tutor but rarely used it for real math talk, and in a young readers' study nearly half of children never logged in. The same 'only about 5% use it as intended' pattern was seen with earlier math software.

Does ChatGPT harm the brain or make kids worse at thinking?

There is no evidence it damages the brain. Small studies in adults and college students suggest heavy reliance lowers effort, memory of one's own work, and self-checking ("metacognitive laziness"). Most of this research is preprints, surveys or small samples, so treat headlines with care.

Do study modes like ChatGPT Study Mode, Gemini Guided Learning or Claude learning mode work?

These modes copy the hint-first design that did well in research, but there are no independent trials of the modes themselves yet. Company descriptions explain how they behave, not whether children learn more. Children can also switch them off.

Which AI tutoring results come from the companies themselves?

Several do: Google's LearnLM papers and the Eedi trial, Anthropic's usage reports, the Rori study in Ghana, and Khan Academy's blog. They can be useful, but independent trials like the University of Toronto's Khanmigo study and Stanford's Tutor CoPilot trial carry more weight.

Can AI help in places without enough teachers?

Early results are encouraging. A teacher-guided GPT-4 program in Nigeria produced gains of about 0.3 standard deviations in six weeks, and a WhatsApp math tutor in Ghana showed similar-sized gains, though the Ghana study was run by the tool's makers. Earlier adaptive software in India also worked well.

What do big reports from Brookings, OECD and UNESCO say?

They are cautious. Brookings' 2026 global review says the risks to children currently outweigh the benefits but the path can still change, and urges tools that teach rather than answer. UNESCO calls for age limits and human-centered design, and the OECD reviews how generative AI is being used in schools.

How should my child use an AI tutor at home to actually learn?

The research points to a few habits: try the problem first, ask the AI for a hint or explanation rather than the answer, and check yourself later without AI. Students who used chatbots to explain did better than those who used them to produce solutions. An adult nearby also helps kids keep using it.

Where can I find the latest research on AI tutoring?

Stanford's SCALE repository collects over 1,600 studies on AI in preK-12 education and is updated monthly. Its 2026 review summarizes the strongest ones. The Hechinger Report's Proof Points column explains new studies in plain language.

The Big Index

AI as a tutor

Can a chatbot teach? The trials, the tutoring systems that came before, and where AI help turns into a crutch.

103 entries
84 free
12 must-reads ★

Questions parents ask

Do AI tutors actually help kids learn?
Sometimes, and usually by a small amount. Well-designed tutors have shown real gains in a few trials (Harvard physics, Nigeria, Ghana), but the biggest independent school trial, of Khanmigo, found small gains about the same as ordinary practice software. Stanford's 2026 review found only 20 strong studies and none with US K-12 students.
- The Evidence Base on AI in K-12: A 2026 Review (Fesler, Martinez, Agnew & Loeb, Stanford SCALE)
- AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting (Kestin, Miller et al., Harvard)
- One Click Away: AI Tutoring with Khanmigo in a Two-Year School Experiment (Oreopoulos & Low, University of Toronto)
- From chalkboards to chatbots: Transforming learning in Nigeria, one prompt at a time (De Simone et al., World Bank)
More in AI as a tutor
Can using ChatGPT for homework make my child learn less?
There is good evidence it can. In a trial with about 1,000 high schoolers, plain ChatGPT raised practice scores but students did 17% worse on a later test without it. Large platform data suggests the same pattern on easy-to-paste math problems. Guardrails that give hints instead of answers reduced most of the harm.
- Generative AI without guardrails can harm learning: Evidence from high school mathematics (Bastani, Bastani, Sungu et al., Penn/Wharton)
- Faster Completion, Less Learning: Generative AI Reduced Study Time on Math Problems and the Knowledge They Build (Rismanchian et al., UC Irvine and McGraw Hill)
- Kids who use ChatGPT as a study assistant do worse on tests (Jill Barshay, The Hechinger Report)
More in AI as a tutor
What makes an AI tutor good rather than harmful?
The studies that went well share a pattern: the AI gives one step or hint at a time, holds back the answer, is fed correct solutions, and steps in after a mistake. Tools built that way look better than general chatbots, though even good design does not help if kids do not use it.
- AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting (Kestin, Miller et al., Harvard)
- Making AI Tutoring Productive: Evidence from a Mastery-Based Math Practice Experiment (Oreopoulos, Liut, Sungu & Low)
- The Evidence Base on AI in K-12: A 2026 Review (Fesler, Martinez, Agnew & Loeb, Stanford SCALE)
- AI Meets the Classroom: When Do Large Language Models Harm Learning? (Lehmann, Cornelius & Sting)
More in AI as a tutor
Is an AI tutor as good as a human tutor?
Not yet as a replacement. Human tutoring several times a week has strong evidence (about a third of a standard deviation on average). AI-only tutoring has much thinner evidence, and Stanford's tutoring center says it does not yet meet that bar. The best-supported use right now is AI helping human tutors.
- The Impressive Effects of Tutoring on PreK-12 Learning: A Systematic Review and Meta-Analysis of the Experimental Evidence (Nickow, Oreopoulos & Quan)
- AI Tutoring is Not a Monolith: What We Actually Know (National Student Support Accelerator, Stanford)
- Tutor CoPilot: A Human-AI Approach for Scaling Real-Time Expertise (Wang, Ribeiro, Robinson, Loeb & Demszky, Stanford)
- AI tutoring can safely and effectively support students: An exploratory RCT in UK classrooms (LearnLM Team, Google & Eedi)
More in AI as a tutor
What is Bloom's "2 sigma problem" and is it true?
In 1984 Benjamin Bloom reported that tutored students did better than 98% of classroom students, a two standard deviation gain. Later reviews found real tutoring effects are much smaller, around 0.3 to 0.8, so promises that AI will deliver two sigma are not supported.
- The 2 Sigma Problem: The Search for Methods of Group Instruction as Effective as One-to-One Tutoring (Benjamin Bloom)
- Two-Sigma Tutoring: Separating Science Fiction from Science Fact (Paul T. von Hippel, Education Next)
- The Relative Effectiveness of Human Tutoring, Intelligent Tutoring Systems, and Other Tutoring Systems (Kurt VanLehn)
Are AI tutors new, or has this been tried before?
Computer tutors have been studied for over 40 years. Older step-by-step tutors like Cognitive Tutor, ASSISTments and AutoTutor had solid, moderate effects in many trials. Chatbots are new in how they talk, but the lessons about feedback, practice and usage carry over.
- Effectiveness of Intelligent Tutoring Systems: A Meta-Analytic Review (Kulik & Fletcher)
- Effectiveness of Cognitive Tutor Algebra I at Scale (Pane, Griffin, McCaffrey & Karam, RAND)
- Online Mathematics Homework Increases Student Achievement (Roschelle, Feng, Murphy & Mason)
- Cognitive Tutors: Lessons Learned (Anderson, Corbett, Koedinger & Pelletier)
Why don't kids use AI tutors much, even good ones?
Usage is the biggest problem found so far. In the Khanmigo trial most students tried the tutor but rarely used it for real math talk, and in a young readers' study nearly half of children never logged in. The same 'only about 5% use it as intended' pattern was seen with earlier math software.
- One Click Away: AI Tutoring with Khanmigo in a Two-Year School Experiment (Oreopoulos & Low, University of Toronto)
- Access is Not Enough: Human Support Improves Engagement with AI Tutoring (Robinson, Gormley, Ribeiro & Loeb, Stanford)
- The 5 Percent Problem (Laurence Holt, Education Next)
- Students didn't get answers from Khanmigo. They didn't want its questions, either (Jill Barshay, The Hechinger Report)
Does ChatGPT harm the brain or make kids worse at thinking?
There is no evidence it damages the brain. Small studies in adults and college students suggest heavy reliance lowers effort, memory of one's own work, and self-checking ("metacognitive laziness"). Most of this research is preprints, surveys or small samples, so treat headlines with care.
- Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task (Kosmyna et al., MIT Media Lab)
- Your Brain on ChatGPT (project site) (Kosmyna et al., MIT Media Lab)
- Beware of Metacognitive Laziness: Effects of Generative Artificial Intelligence on Learning Motivation, Processes, and Performance (Fan, Tang ... Gašević)
- Against "Brain Damage" (Ethan Mollick, One Useful Thing)
Do study modes like ChatGPT Study Mode, Gemini Guided Learning or Claude learning mode work?
These modes copy the hint-first design that did well in research, but there are no independent trials of the modes themselves yet. Company descriptions explain how they behave, not whether children learn more. Children can also switch them off.
- OpenAI launches Study Mode in ChatGPT (TechCrunch)
- Guided Learning in Gemini: From answers to understanding (Google)
- Introducing Claude for Education (Anthropic)
- Generative AI without guardrails can harm learning: Evidence from high school mathematics (Bastani, Bastani, Sungu et al., Penn/Wharton)
Which AI tutoring results come from the companies themselves?
Several do: Google's LearnLM papers and the Eedi trial, Anthropic's usage reports, the Rori study in Ghana, and Khan Academy's blog. They can be useful, but independent trials like the University of Toronto's Khanmigo study and Stanford's Tutor CoPilot trial carry more weight.
- AI tutoring can safely and effectively support students: An exploratory RCT in UK classrooms (LearnLM Team, Google & Eedi)
- LearnLM: Improving Gemini for Learning (LearnLM Team, Google)
- Effective and Scalable Math Support: Evidence on the Impact of an AI-Tutor on Math Achievement in Ghana (Henkel et al., Rising Academies)
- One Click Away: AI Tutoring with Khanmigo in a Two-Year School Experiment (Oreopoulos & Low, University of Toronto)
Can AI help in places without enough teachers?
Early results are encouraging. A teacher-guided GPT-4 program in Nigeria produced gains of about 0.3 standard deviations in six weeks, and a WhatsApp math tutor in Ghana showed similar-sized gains, though the Ghana study was run by the tool's makers. Earlier adaptive software in India also worked well.
- From chalkboards to chatbots: Transforming learning in Nigeria, one prompt at a time (De Simone et al., World Bank)
- Effective and Scalable Math Support: Evidence on the Impact of an AI-Tutor on Math Achievement in Ghana (Henkel et al., Rising Academies)
- Disrupting Education? Experimental Evidence on Technology-Aided Instruction in India (Muralidharan, Singh & Ganimian)
What do big reports from Brookings, OECD and UNESCO say?
They are cautious. Brookings' 2026 global review says the risks to children currently outweigh the benefits but the path can still change, and urges tools that teach rather than answer. UNESCO calls for age limits and human-centered design, and the OECD reviews how generative AI is being used in schools.
- A new direction for students in an AI world: Prosper, prepare, protect (Burns, Winthrop et al., Brookings)
- Guidance for generative AI in education and research (UNESCO)
- OECD Digital Education Outlook 2026 (OECD)
How should my child use an AI tutor at home to actually learn?
The research points to a few habits: try the problem first, ask the AI for a hint or explanation rather than the answer, and check yourself later without AI. Students who used chatbots to explain did better than those who used them to produce solutions. An adult nearby also helps kids keep using it.
- AI Meets the Classroom: When Do Large Language Models Harm Learning? (Lehmann, Cornelius & Sting)
- Against "Brain Damage" (Ethan Mollick, One Useful Thing)
- Access is Not Enough: Human Support Improves Engagement with AI Tutoring (Robinson, Gormley, Ribeiro & Loeb, Stanford)
- Slow math: Kids may learn more when AI makes them review mistakes (Barshay & Fasiang, Hechinger Report)
Where can I find the latest research on AI tutoring?
Stanford's SCALE repository collects over 1,600 studies on AI in preK-12 education and is updated monthly. Its 2026 review summarizes the strongest ones. The Hechinger Report's Proof Points column explains new studies in plain language.
- AI Hub for Education Research Repository (Stanford SCALE Initiative)
- The Evidence Base on AI in K-12: A 2026 Review (Fesler, Martinez, Agnew & Loeb, Stanford SCALE)
- National Student Support Accelerator (Stanford University)

Must-reads

StudyFree
AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting
Kestin, Miller et al., Harvard, 2025
194 Harvard physics students each tried a custom AI tutor at home and an active-learning class. They learned about twice as much in less time with the tutor, which gave one step at a time and never the full answer. College students, two topics.
nature.com
StudyFree
Generative AI without guardrails can harm learning: Evidence from high school mathematics
Bastani, Bastani, Sungu et al., Penn/Wharton, 2025
Field experiment with nearly 1,000 Turkish high schoolers. Plain ChatGPT boosted practice scores but students then did 17% worse on an exam without it. A hint-giving tutor version avoided most of that harm. The key study on AI as a crutch.
pnas.org
BlogFree
From chalkboards to chatbots: Transforming learning in Nigeria, one prompt at a time
De Simone et al., World Bank, 2025
A six-week after-school program in Nigeria where teachers guided students using a GPT-4-based tutor for English. Gains were about 0.3 standard deviations, beating most education programs in poor countries. Girls gained more. Teachers were in the room.
blogs.worldbank.org
StudyFree
One Click Away: AI Tutoring with Khanmigo in a Two-Year School Experiment
Oreopoulos & Low, University of Toronto, 2026
Independent two-year randomized trial in 18 Tennessee middle schools. Khanmigo gave small gains, about the same as plain Khan Academy practice. Most students tried it but rarely used it for real math talk. Working paper, not yet peer reviewed.
nber.org
StudyFree
AI tutoring can safely and effectively support students: An exploratory RCT in UK classrooms
LearnLM Team, Google & Eedi, 2025
165 UK secondary students got math chat tutoring from humans or from Google's LearnLM, with expert tutors approving each AI message. AI-supported students did at least as well and solved new problems slightly more often. Small, company-run study.
storage.googleapis.com
StudyFree
Tutor CoPilot: A Human-AI Approach for Scaling Real-Time Expertise
Wang, Ribeiro, Robinson, Loeb & Demszky, Stanford, 2024
Randomized trial where AI whispered suggestions to human tutors during live math sessions. Students were 4 points more likely to master topics, 9 points with weaker tutors, at about $20 per tutor per year. Tutors gave away fewer answers.
arxiv.org
StudyPaid
The 2 Sigma Problem: The Search for Methods of Group Instruction as Effective as One-to-One Tutoring
Benjamin Bloom, 1984
The classic paper claiming tutored students beat 98% of classroom peers. Almost every AI tutor pitch cites it. The two-sigma figure came from small studies and has not held up at that size; see von Hippel's critique.
doi.org
ArticleFree
Two-Sigma Tutoring: Separating Science Fiction from Science Fact
Paul T. von Hippel, Education Next, 2024
Shows Bloom's famous result rested on narrow tests and extra ingredients. Realistic tutoring effects are about a third of a standard deviation, a fairer target for AI tutors. Best single read on what to expect.
educationnext.org
Meta-analysisFree
The Impressive Effects of Tutoring on PreK-12 Learning: A Systematic Review and Meta-Analysis of the Experimental Evidence
Nickow, Oreopoulos & Quan, 2020
Pools nearly 100 randomized trials of human tutoring. Average gain about 0.37 standard deviations, larger with teachers or trained paraprofessionals and in early grades. The benchmark any AI tutor should be judged against.
nber.org
ReportFree
AI Tutoring is Not a Monolith: What We Actually Know
National Student Support Accelerator, Stanford, 2026
Sorts AI tutoring into types by how much human contact each keeps. Bottom line: AI that boosts human tutors has evidence; AI-only tutoring is useful practice but not yet proven tutoring, and depends on adults setting it up.
nssa.stanford.edu
ReviewFree
The Evidence Base on AI in K-12: A 2026 Review
Fesler, Martinez, Agnew & Loeb, Stanford SCALE, 2026
Of 800+ papers, only 20 gave strong causal evidence, none with US K-12 students. Pattern: AI boosts work while kids have it, gains often fade without it, and tools that guide rather than answer look best. The best current summary.
scale.stanford.edu
ReportFree
A new direction for students in an AI world: Prosper, prepare, protect
Burns, Winthrop et al., Brookings, 2026
A year-long global review with 500+ people in 50 countries and 400+ studies. It concludes that for now AI's risks to children's learning outweigh its benefits, and gives 12 steps, including picking tools that teach instead of answer.
brookings.edu

All 103

AI tutor trials

★AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational settingKestin, Miller et al., Harvard · 2025 · Study · Free194 Harvard physics students each tried a custom AI tutor at home and an active-learning class. They learned about twice as much in less time with the tutor, which gave one step at a time and never the full answer. College students, two topics.
Professor tailored AI tutor to physics course. Engagement doubled.Harvard Gazette · 2024 · Article · FreePlain-language story about the Harvard physics AI tutor trial. Useful for the design details: the tutor was fed the instructor's own solutions and told to guide thinking, and the lead author says it should prepare students for class, not replace it.
An AI tutor helped Harvard students learn more physics in less timeJill Barshay, Hechinger Report · 2024 · Article · FreeA careful reporter's take on the Harvard tutor study, with the caveats: under 200 students, sessions were scheduled and supervised, and it is unclear whether students would use such a tool on their own.
★Generative AI without guardrails can harm learning: Evidence from high school mathematicsBastani, Bastani, Sungu et al., Penn/Wharton · 2025 · Study · Free · TeensField experiment with nearly 1,000 Turkish high schoolers. Plain ChatGPT boosted practice scores but students then did 17% worse on an exam without it. A hint-giving tutor version avoided most of that harm. The key study on AI as a crutch.
Without Guardrails, Generative AI Can Harm EducationKnowledge at Wharton · 2024 · Article · FreeShort summary of the Bastani study for non-researchers. Highlights that students who used AI felt more confident than their results justified, and that a tutor built with teacher input avoided the drop in exam scores.
★From chalkboards to chatbots: Transforming learning in Nigeria, one prompt at a timeDe Simone et al., World Bank · 2025 · Blog · Free · TeensA six-week after-school program in Nigeria where teachers guided students using a GPT-4-based tutor for English. Gains were about 0.3 standard deviations, beating most education programs in poor countries. Girls gained more. Teachers were in the room.
From Chalkboards to Chatbots: Evaluating the Impact of Generative AI on Learning Outcomes in NigeriaDe Simone, Tiberti et al., World Bank · 2025 · Study · Free · TeensThe full World Bank working paper behind the Nigeria pilot, with methods and cost-effectiveness. A randomized trial, but a working paper rather than a peer-reviewed journal article.
From chalkboards to chatbots in Nigeria: 7 lessons to pioneer generative AI for educationWorld Bank · 2025 · Blog · FreePractical lessons from running the Nigeria AI tutoring pilot: what the program looked like day to day, why teacher guidance mattered, and what to watch when copying it.
AI/DD Brown Bag Lunch: From Chalkboards to ChatbotsWorld Bank · 2025 · Video · FreeRecorded talk where the World Bank researchers walk through the Nigeria AI tutoring results. Good if you prefer to hear the study explained.
Effective and Scalable Math Support: Evidence on the Impact of an AI-Tutor on Math Achievement in GhanaHenkel et al., Rising Academies · 2024 · Study · Free · Ages 8–12About 1,000 students in grades 3-9 in Ghana used Rori, a math tutor on WhatsApp, twice a week for eight months. Math growth was clearly higher than in control schools. Preprint, first year only, written by the team behind the tool.
★One Click Away: AI Tutoring with Khanmigo in a Two-Year School ExperimentOreopoulos & Low, University of Toronto · 2026 · Study · Free · Ages 8–12Independent two-year randomized trial in 18 Tennessee middle schools. Khanmigo gave small gains, about the same as plain Khan Academy practice. Most students tried it but rarely used it for real math talk. Working paper, not yet peer reviewed.
What I found compelling in a new randomized trial of Khan Academy in math interventionSal Khan, Khan Academy · 2026 · Blog · Free · Ages 8–12Khan Academy's founder responds to the independent Khanmigo trial. He stresses the larger gain for students who stayed all year and describes product changes. Written by the company that makes the product.
Making AI Tutoring Productive: Evidence from a Mastery-Based Math Practice ExperimentOreopoulos, Liut, Sungu & Low · 2026 · Study · Free · Ages 8–12Over 6,000 middle schoolers practiced fractions with or without an AI tutor, with or without a three-in-a-row mastery rule. AI helped most right after mistakes, and AI plus mastery did best, but gains were small and short-term. Working paper.
Slow math: Kids may learn more when AI makes them review mistakesBarshay & Fasiang, Hechinger Report · 2026 · Article · Free · Ages 8–12Readable summary of the AI-plus-mastery fractions experiment. The gain was about 3 percentage points and did not reach harder problems, so the lesson is modest: AI that walks through errors beats AI that shows answers.
Access is Not Enough: Human Support Improves Engagement with AI TutoringRobinson, Gormley, Ribeiro & Loeb, Stanford · 2026 · Study · Free · Ages 5–8Two trials with 355 children in grades 1-5 using an AI reading platform. Nearly half of kids left alone never used it. Sitting with a human tutor raised use, but use stayed low and reading did not improve. Working paper.
★AI tutoring can safely and effectively support students: An exploratory RCT in UK classroomsLearnLM Team, Google & Eedi · 2025 · Study · Free · Teens165 UK secondary students got math chat tutoring from humans or from Google's LearnLM, with expert tutors approving each AI message. AI-supported students did at least as well and solved new problems slightly more often. Small, company-run study.
Just launched: our second AI tutor RCTEedi · 2026 · Blog · Free · TeensEedi describes its larger follow-up trial with Google DeepMind: 1,525 UK students, four groups including human-only tutoring, and an AI tutor that only switches on after a specific wrong answer. Company post; results pending.
AI Tutoring Enhances Student Learning Without Crowding Out Reading EffortFischer, Rau & Rilke, IZA · 2025 · Study · FreeRandomized experiment with 334 university students. Access to an AI tutor raised test scores by about 0.23 standard deviations over textbook-only study, helping most those who started with less knowledge. Discussion paper, college age.
Generative AI alone may not be enough: Evaluating AI Support for Learning Mathematical ProofChen ... Koedinger, Carnegie Mellon · 2025 · Study · FreeIn a college proof course, an AI tutor improved homework but not exam scores. Students with low confidence leaned on the chatbot more and did worse, while a structured proof-review tool predicted better finals. Preprint.
From Cognitive Relief to Affective Engagement: An Empirical Comparison of AI Chatbots and Instructional Scaffolding in Physics EducationBecker et al. · 2025 · Study · Free · Teens273 German ninth graders solved a physics problem with a chatbot, tiered hints, or text. The chatbot and the hints both lightened mental load; the chatbot felt better to students but did not beat simple hints by a clear margin. Preprint.
The Impact of Large Language Models on Students: A Randomised Study of Socratic vs. Non-Socratic AI and the Role of Step-by-Step ReasoningBlasco & Charisi · 2024 · Study · Free · TeensA school-age randomized study comparing a chatbot that asks guiding questions with one that just explains. Counted by Stanford's 2026 review among the few strong causal studies of AI with K-12 students. Working paper on SSRN.
ChatGPT-generated help produces learning gains equivalent to human tutor-authored help on mathematics skillsPardos & Bhandari, UC Berkeley · 2024 · Study · FreeCompared math hints written by ChatGPT with hints written by human tutors in an online system. Learning gains were similar, which suggests AI can cheaply draft help content, though answers still need checking. Peer reviewed, open access.
AI Meets the Classroom: When Do Large Language Models Harm Learning?Lehmann, Cornelius & Sting · 2024 · Study · FreeLab and field experiments showing that how students use a chatbot matters most. Asking it to explain helps understanding; asking it to produce solutions lets students cover more ground but understand less. It also widened gaps between strong and weak students. Preprint.
Learning To Prompt: Improving Student Engagement With Adaptive LLM-Based High-School TutoringChang, Hogan, Plaat & van der Meer · 2026 · Study · Free · TeensA/B test with 359 high schoolers of a tutor that switches teaching style by subject. It shortened chats while keeping quality and got more students to try practice exercises. Engagement result, not a learning-gain result. Preprint.
Effects of LLM use and note-taking on reading comprehension and memory: A randomised experiment in secondary schoolsKreijkes et al., Cambridge and Microsoft Research · 2026 · Study · Paid · TeensRandomized school experiment testing whether reading with a chatbot changes what teens understand and remember compared with taking notes. Valuable because it checks memory later, not just in the moment. Peer reviewed; partly industry researchers.

AI helping human tutors

★Tutor CoPilot: A Human-AI Approach for Scaling Real-Time ExpertiseWang, Ribeiro, Robinson, Loeb & Demszky, Stanford · 2024 · Study · Free · Ages 8–12Randomized trial where AI whispered suggestions to human tutors during live math sessions. Students were 4 points more likely to master topics, 9 points with weaker tutors, at about $20 per tutor per year. Tutors gave away fewer answers.
Automated feedback improves teachers' questioning quality in brick-and-mortar classrooms: Opportunities for further enhancementDemszky, Liu, Hill et al. · 2025 · Study · PaidTeachers who got automatic AI feedback on their own classroom talk asked better questions. Shows another use of AI: coaching the adult rather than replacing them. Peer reviewed.
Home-Tutoring Services Assisted with Technology: Investigating the Role of Artificial Intelligence Using a Randomized Field ExperimentKim, Kim, Kwak & Lee · 2021 · Study · PaidA field experiment with a home-tutoring company testing what happens when AI tools support human tutors. An early example, before chatbots, of AI and people working together on tutoring. Paywalled journal.

Offloading and over-reliance

Your Brain on ChatGPT (project site)Kosmyna et al., MIT Media Lab · 2025 · Article · FreeThe researchers' own site for the study, with summaries and an FAQ. Worth reading before trusting headlines that claim ChatGPT "damages" brains, which the study does not show.
The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge WorkersLee, Sarkar et al., Microsoft Research and CMU · 2025 · Study · FreeSurvey of 319 office workers. The more people trusted AI, the less critical thinking they reported; the more they trusted themselves, the more. Adults, self-report, industry researchers, but a clear picture of how trust shapes thinking.
Beware of Metacognitive Laziness: Effects of Generative Artificial Intelligence on Learning Motivation, Processes, and PerformanceFan, Tang ... Gašević · 2024 · Study · Free117 university students wrote with help from ChatGPT, a human expert, other tools, or nothing. ChatGPT raised essay scores but not real knowledge gain, and students planned and checked their own work less. Coined "metacognitive laziness." Also in BJET.
AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical ThinkingMichael Gerlich · 2025 · Study · FreeSurvey and interviews with 666 people linking heavy AI use to lower critical-thinking scores, strongest in younger users. Widely shared, but it is correlational and self-reported, so it cannot show that AI causes the drop.
Cognitive ease at a cost: LLMs reduce mental effort but compromise depth in student scientific inquiryStadler, Bannert & Sailer · 2024 · Study · PaidUniversity students researching a science question with ChatGPT felt less mental strain than those using Google, but wrote shallower, weaker arguments. Easier is not the same as learning more.
Faster Completion, Less Learning: Generative AI Reduced Study Time on Math Problems and the Knowledge They BuildRismanchian et al., UC Irvine and McGraw Hill · 2026 · Study · Free · TeensMillions of records from the ALEKS math platform. After ChatGPT arrived, students sped through easy-to-paste word problems and then scored much worse on those on proctored tests. Indirect evidence, preprint, co-written by the platform's maker.
Faster solutions, lower test scores: How AI is eroding math skillsJill Barshay, Hechinger Report · 2026 · Article · Free · TeensPlain explanation of the ALEKS data study: word-problem accuracy on supervised tests fell from about 80% to about 60% while graphing problems, which are harder to paste into a chatbot, held steady.
Generative AI Can Harm TeachingSungu, Duckworth et al., Penn · 2026 · Study · FreeRandomized trial with 193 teachers in Turkey. Teachers given a ChatGPT lesson assistant made blander lessons, and students of weaker teachers lost ground. A reminder that AI shortcuts can hit kids through adults too. Unpublished draft.
Teachers save time with AI. Their students may pay the priceJill Barshay, Hechinger Report · 2026 · Article · FreeSummary of the teacher-assistant trial in Turkey: students rated AI-assisted lessons as less interesting, and the harm landed mostly in classes of weaker teachers.

Company research and features

How Google's LearnLM generative AI models support teachers and learnersGoogle · 2024 · Blog · FreeGoogle's launch post for LearnLM, its tutoring-tuned Gemini models, built around five learning-science ideas such as active learning and managing mental load. Company announcement, not evidence of results.
LearnLM: Improving Gemini for LearningLearnLM Team, Google · 2024 · Report · FreeGoogle's technical report on training Gemini to follow teaching instructions. Expert raters preferred it to other chatbots for tutoring. Rater preference is not the same as students learning more. Company research.
Towards Responsible Development of Generative AI for Education: An Evaluation-Driven ApproachJurenka, Kunesch et al., Google DeepMind · 2024 · Report · FreeGoogle DeepMind's long paper on how to measure whether an AI tutor teaches well, turning learning science into seven tests. Useful for anyone asking what "good tutoring" means for a chatbot. Company research.
OpenAI launches Study Mode in ChatGPTTechCrunch · 2025 · Article · FreeNews report on ChatGPT's study mode, which quizzes students and gives hints rather than answers. Built with custom instructions from teachers and learning experts, but students can switch it off in one click.
Introducing Claude for EducationAnthropic · 2025 · Blog · FreeAnthropic's launch of Claude for universities, including a learning mode that asks students how they would approach a problem instead of solving it. Company announcement; aimed at college, not children.
Anthropic Education Report: How University Students Use ClaudeAnthropic · 2025 · Report · FreeAnalysis of about a million anonymized student chats. Many asked Claude to create or analyze for them, which the company itself flags as a sign students may hand off thinking. Company research on its own product.
Anthropic Education Report: How Educators Use ClaudeAnthropic · 2025 · Report · FreeStudy of 74,000 professor chats: most built lessons and teaching tools, but nearly half of grading chats were fully handed to AI even though faculty rated that use worst. Company research.
Introducing Claude for TeachersAnthropic · 2026 · Blog · FreeAnthropic's free offer for verified US K-12 teachers, linking Claude to standards, open curricula and tools like ASSISTments and Eedi, with a Detroit pilot to study effects. Company announcement.
KhanmigoKhan Academy · 2023 · App · Free + paid · Ages 5–8Khan Academy's AI tutor and teacher assistant, designed to ask guiding questions rather than give answers, built into Khan Academy practice. Nonprofit maker; independent trials so far show small gains similar to Khan Academy alone.

Intelligent tutoring history

★The 2 Sigma Problem: The Search for Methods of Group Instruction as Effective as One-to-One TutoringBenjamin Bloom · 1984 · Study · PaidThe classic paper claiming tutored students beat 98% of classroom peers. Almost every AI tutor pitch cites it. The two-sigma figure came from small studies and has not held up at that size; see von Hippel's critique.
★Two-Sigma Tutoring: Separating Science Fiction from Science FactPaul T. von Hippel, Education Next · 2024 · Article · FreeShows Bloom's famous result rested on narrow tests and extra ingredients. Realistic tutoring effects are about a third of a standard deviation, a fairer target for AI tutors. Best single read on what to expect.
The Relative Effectiveness of Human Tutoring, Intelligent Tutoring Systems, and Other Tutoring SystemsKurt VanLehn · 2011 · Review · PaidLandmark review finding human tutoring was far less than two sigma (about 0.79) and step-by-step computer tutors came close (about 0.76). Suggests the step-level feedback, not the human, does much of the work.
Effectiveness of Intelligent Tutoring Systems: A Meta-Analytic ReviewKulik & Fletcher · 2016 · Meta-analysis · PaidPooled 50 controlled studies of pre-chatbot tutoring software and found sizable average gains, larger on tests the researchers designed than on standardized tests. A good baseline for judging new AI claims.
Intelligent tutoring systems and learning outcomes: A meta-analysisMa, Adesope, Nesbit & Liu · 2014 · Meta-analysis · PaidReview of 107 effect sizes: computer tutors beat teacher-led large-group teaching and plain computer drills, and roughly matched small-group or one-to-one human tutoring.
A meta-analysis of the effectiveness of intelligent tutoring systems on college students' academic learningSteenbergen-Hu & Cooper · 2014 · Meta-analysis · PaidFound moderate average benefits of tutoring software for college students compared with typical instruction. A useful check on whether the older software already delivered what AI tutors now promise.
Cognitive Tutors: Lessons LearnedAnderson, Corbett, Koedinger & Pelletier · 1995 · Article · PaidThe founders of the Carnegie Mellon Cognitive Tutor explain a decade of building software that tracks each step of a student's work. The roots of today's math tutors like MATHia.
Effectiveness of Cognitive Tutor Algebra I at ScalePane, Griffin, McCaffrey & Karam, RAND · 2014 · Study · Paid · TeensA large RAND randomized trial in 147 schools. Cognitive Tutor Algebra showed no effect in year one and a meaningful gain in year two, a lesson that new tools often need time and teacher practice to work.
Online Mathematics Homework Increases Student AchievementRoschelle, Feng, Murphy & Mason · 2016 · Study · Free · Ages 8–12Randomized trial in Maine middle schools of ASSISTments, free homework software giving instant feedback and teacher reports. Students gained notably, with the biggest gains for lower achievers. Open access.
AutoTutor: A tutor with dialogue in natural languageGraesser et al., University of Memphis · 2004 · Article · PaidDescribes AutoTutor, a pre-ChatGPT tutor that held conversations in plain English, giving hints and prompts like a human tutor. Shows that "chat with a tutor" research is decades old.
Disrupting Education? Experimental Evidence on Technology-Aided Instruction in IndiaMuralidharan, Singh & Ganimian · 2019 · Study · Paid · Ages 8–12Randomized trial of Mindspark, adaptive software used in after-school centers in Delhi. Middle schoolers made large gains in math and Hindi because the software met them at their real level, often years behind grade.
Education Technology: An Evidence-Based ReviewEscueta, Quan, Nickow & Oreopoulos · 2017 · Review · FreeBroad review of randomized studies on ed tech. Computer-assisted learning, especially in math, was among the most consistently helpful; simply giving kids devices was not. Later published in the Journal of Economic Literature.
A systematic review of AI-driven intelligent tutoring systems (ITS) in K-12 educationLétourneau et al. · 2025 · Review · FreeRecent open-access review of tutoring software studies with school-age children. Effects are generally positive but shrink when compared with other good teaching, and many studies are small. Good for a sober overview.
Evolution and Revolution in Artificial Intelligence in EducationRoll & Wylie · 2016 · Review · PaidA look back at 25 years of AI in education research, arguing the field should move beyond drill software toward collaboration and real-world settings. Helpful context for how today's chatbots fit a longer story.

Reviews and reports

Does ChatGPT enhance student learning? A systematic review and meta-analysis of experimental studiesDeng, Jiang, Yu, Lu & Liu · 2025 · Meta-analysis · PaidPools experiments on ChatGPT and learning. Average effects on performance look positive, but most studies are short, small and measure work done with AI present, so durable learning is less clear.
RETRACTED: The effect of ChatGPT on students' learning performance, learning perception, and higher-order thinking: insights from a meta-analysisWang & Fan · 2025 · Meta-analysis · FreeA widely shared 2025 meta-analysis reporting large ChatGPT benefits that has since been retracted. Listed here as a warning: if an article or chatbot cites it as proof, treat the claim as unsupported.
★The Evidence Base on AI in K-12: A 2026 ReviewFesler, Martinez, Agnew & Loeb, Stanford SCALE · 2026 · Review · FreeOf 800+ papers, only 20 gave strong causal evidence, none with US K-12 students. Pattern: AI boosts work while kids have it, gains often fade without it, and tools that guide rather than answer look best. The best current summary.
AI Hub for Education Research RepositoryStanford SCALE Initiative · 2025 · Data · FreeSearchable, monthly-updated collection of over 1,600 studies on AI in preK-12 education, filterable by age, use and study design. Most entries are preprints, so check study quality.
★A new direction for students in an AI world: Prosper, prepare, protectBurns, Winthrop et al., Brookings · 2026 · Report · FreeA year-long global review with 500+ people in 50 countries and 400+ studies. It concludes that for now AI's risks to children's learning outweigh its benefits, and gives 12 steps, including picking tools that teach instead of answer.
OECD Digital Education Outlook 2026OECD · 2026 · Report · FreeThe OECD's flagship report on digital education, focused on generative AI in learning and teaching, with evidence reviews and country examples. Free to read online.
OECD Digital Education Outlook 2023OECD · 2023 · Report · FreeEarlier OECD edition on building effective digital education systems, including adaptive learning and data use. Helpful background before the 2026 generative AI edition.
Guidance for generative AI in education and researchUNESCO · 2023 · Guide · FreeThe first global guidance on generative AI in schools. It calls for privacy rules, teacher training and a minimum age for using chatbots on one's own, and for human-centered, age-appropriate design.
Artificial intelligence in educationUNESCO · Organization · FreeUNESCO's hub for its work on AI in education, including AI competency frameworks for students and teachers, policy guides for governments, and links to its global guidance.
S.A.F.E. by Design: Policy, Research, and Practice Recommendations for AI Companions in EducationEDSAFE AI Alliance · 2026 · Guide · FreeTask force report on chatbots that act as tutors or companions for students, with safety and design recommendations for schools and companies.
AI Tutors: Hype or Hope for Education?John Bailey & John Warner, Education Next · 2025 · Article · FreeA debate. Bailey argues on-demand AI expertise could help students now; Warner argues AI tutors repeat a century of failed teaching machines and miss why kids disengage. Good for seeing both sides.
The 5 Percent ProblemLaurence Holt, Education Next · 2024 · Article · FreeOnline math programs work for kids who use them as recommended, but only about 5% do, and they are often the kids already ahead. The same usage problem now shows up in AI tutor trials.

Tools and organizations

MATHiaCarnegie Learning · App · Paid · TeensThe current version of the Carnegie Mellon Cognitive Tutor for grades 6-12 math, sold to schools. It follows each step of a student's work and gives targeted hints. Commercial product with a long research record.
ASSISTmentsASSISTments (nonprofit) · Tool · Free · Ages 8–12Free math homework tool for teachers that gives students instant feedback and shows teachers who is stuck. Backed by a randomized trial and widely used as a research platform.
LearnLabCarnegie Mellon Simon Initiative · Organization · FreeCarnegie Mellon's learning-science research center, behind much of the intelligent tutoring research of the last 30 years. It shares tools, datasets and study results with other researchers.
RoriRising Academies · App · Free · Ages 8–12A math tutor that works over WhatsApp on basic phones, built for students in Africa. Tested in a randomized trial in Ghana run by its own team.
EediEedi · Tool · Free + paid · Ages 8–12UK math platform built around diagnostic questions that reveal each student's misconception. Its AI tutor only steps in after a specific wrong answer, and it runs trials with Google DeepMind.
Digital PromiseDigital Promise · Organization · FreeNonprofit that studies education technology and certifies products that show research-based design, including AI tools. Useful when checking claims an app makes about itself.
Center on Reinventing Public EducationCRPE, Arizona State University · Organization · FreeArizona State research center that tracks how schools adopt AI, with reports on early-adopter districts, AI tutoring pilots and what school leaders are learning.
Teaching Systems LabMIT · Organization · FreeJustin Reich's MIT lab studying how new technology, including generative AI, actually changes teaching in real schools. Home of the TeachLab podcast and its Homework Machine series.

Human tutoring evidence

★The Impressive Effects of Tutoring on PreK-12 Learning: A Systematic Review and Meta-Analysis of the Experimental EvidenceNickow, Oreopoulos & Quan · 2020 · Meta-analysis · FreePools nearly 100 randomized trials of human tutoring. Average gain about 0.37 standard deviations, larger with teachers or trained paraprofessionals and in early grades. The benchmark any AI tutor should be judged against.
Not Too Late: Improving Academic Outcomes Among AdolescentsGuryan, Ludwig et al. · 2021 · Study · Free · TeensTwo big Chicago trials of daily in-school math tutoring by Saga Education. Teens gained 0.16 to 0.37 standard deviations in math and got better grades, showing it is never too late to catch up. Published in the AER in 2023.
National Student Support AcceleratorStanford University · Organization · FreeStanford center that studies and supports high-impact tutoring, with toolkits, a research database, and recent briefs on what AI tutoring can and cannot do yet.
What is High-Impact Tutoring?National Student Support Accelerator · Guide · FreeShort explainer of the tutoring recipe that works: sessions several times a week, small groups, the same trained tutor, and good materials. A checklist parents can use to judge any tutoring service, human or AI.
Toolkit for Tutoring ProgramsNational Student Support Accelerator · Guide · FreeFree step-by-step toolkit for schools and community groups starting tutoring programs, covering tutor hiring and training, scheduling, materials and how to measure whether it is working.
Design Principles for Accelerating Student Learning With High-Impact TutoringRobinson, Kraft, Loeb & Schueler · 2021 · Guide · FreeFour-page research brief on what makes tutoring work: frequency, small groups, consistent tutors, and links to class content. Clear and practical.
A Blueprint for Scaling Tutoring Across Public SchoolsKraft & Falken · 2020 · Report · FreeThought experiment on how the US could give every student tutoring, covering cost, staffing and using technology to stretch human tutors. Useful context for why people hope AI can help.
★AI Tutoring is Not a Monolith: What We Actually KnowNational Student Support Accelerator, Stanford · 2026 · Report · FreeSorts AI tutoring into types by how much human contact each keeps. Bottom line: AI that boosts human tutors has evidence; AI-only tutoring is useful practice but not yet proven tutoring, and depends on adults setting it up.
AI Tutors Not Yet a Replacement for Humans, Research SaysLinda Jacobson, The 74 · 2026 · Article · FreeNews story with a parent's view of her kids' frustration with an AI reading tutor, and researchers noting most children do not use such tools long enough to benefit.
Saga EducationSaga Education · Organization · Free · TeensNonprofit running daily in-school math tutoring, the program behind the Chicago "Not Too Late" trials, now testing technology to stretch its human tutors.
How High-Impact Tutoring Is Reshaping Post-Pandemic Learning RecoveryHarvard EdCast · 2025 · Podcast · FreeInterview with author Liz Cohen on how schools used tutoring after the pandemic and what made it work. Human tutoring only, but the background every AI tutor debate needs.

Explainers, blogs and podcasts

The Homework ApocalypseEthan Mollick, One Useful Thing · 2023 · Blog · FreeWharton professor's early warning that AI can do nearly all homework, and ideas for how teaching might change, including AI as tutor.
Post-apocalyptic educationEthan Mollick, One Useful Thing · 2024 · Blog · FreeMollick's follow-up: students feel AI help is teaching them when it often is not, and teachers overestimate their ability to spot AI work. Argues for redesigning classes around AI that prompts thinking.
Against "Brain Damage"Ethan Mollick, One Useful Thing · 2025 · Blog · FreeA calm response to scary headlines about AI and the brain. AI does not damage brains, but lazy use can weaken thinking; his advice is to think first, then use AI to push further.
The future of education in a world of AIEthan Mollick, One Useful Thing · 2023 · Blog · FreeMollick's hopeful sketch of classrooms where AI tutors give each student individual help outside class and teachers use class time for discussion and active learning.
Assigning AI: Seven Ways of Using AI in ClassEthan Mollick, One Useful Thing · 2023 · Blog · FreeSeven roles AI can play for students, including tutor, coach and teammate, with sample prompts and what can go wrong with each. Practical for parents trying a tutor prompt at home.
Co-Intelligence: Living and Working with AIEthan Mollick · 2024 · Book · PaidBestselling general book on working with AI, with a chapter on AI as tutor and why effortful learning still matters.
Six Important Findings from the AI in Action Learning TourDan Meyer, Mathworlds · 2026 · Blog · FreeA math educator summarizes classroom visits watching AI tools in use. Tools that respect the teacher did better; many tools were dreary; students copied Google's AI summaries without reading.
MathworldsDan Meyer · Newsletter · FreeNewsletter by a well-known math teacher and ed tech critic, often skeptical of AI tutor claims and focused on what makes kids want to think.
The Homework Machine Ep 1: "Buckle Up, Here it Comes!"Justin Reich, TeachLab (MIT) · 2025 · Podcast · FreeFirst episode of an MIT podcast series on how generative AI arrived in schools uninvited, told through teachers and students. Thoughtful and grounded.
Teaching Students to Think Critically About AIHarvard EdCast · 2025 · Podcast · FreeTwo education professors on teaching kids that AI is not neutral, with ideas like auditing who built a tool and why.
Class Disrupted PodcastThe 74 · Podcast · FreePodcast from the nonprofit education news site The 74 about rethinking how school works, with recurring episodes on AI tutors, personalized learning and new school models.
Artificial Intelligence (topic page)The 74 · Article · FreeRunning collection of education news on AI in schools from a nonprofit newsroom, including AI tutor trials and district rollouts.