AI Tutors vs. Human Tutors: What the Research Shows
September 11, 2026
If a free AI app can walk your child through a fraction problem at 9 pm, why would anyone pay a human tutor $50 an hour to do the same thing? That's the pitch behind nearly every AI tutoring product launched this year, and it deserves a straight answer.
Here it is. Research through 2024 shows human tutoring produces reliable, moderate learning gains, while a growing body of 2025 and 2026 studies shows AI tutoring alone runs into a different problem entirely: kids don't use it consistently unless someone is holding them accountable. The strongest results in the actual research come from combining the two, not picking a side.
That's a more useful answer than either side's marketing, and it's worth understanding why.
What Do People Actually Mean by "AI Tutor"?
This is where most comparisons go wrong before they even start. "AI tutor" gets used to describe at least three different products, and they don't perform the same way in the research.
The first is a conversational chatbot bolted onto an existing platform, like Khan Academy's Khanmigo, that answers questions and walks through problems on demand. The second is a standalone homework help app your child opens independently, with no teacher or parent checking in at all. The third is an AI-copilot that works alongside a live human teacher during a real class, flagging in real time where a student is stuck so the teacher can adjust, which behaves very differently from a chatbot a child talks to alone.
Adaptive practice platforms like IXL sit in a related but separate category. They personalize the sequence of problems using an algorithm, but they aren't conversational, and most parents use them as a supplement to instruction rather than a replacement for it. When a headline claims "AI tutoring works" or "AI tutoring doesn't work," it's worth checking which of these three it actually tested. The research below breaks down cleanly along that line.
Here it is. Research through 2024 shows human tutoring produces reliable, moderate learning gains, while a growing body of 2025 and 2026 studies shows AI tutoring alone runs into a different problem entirely: kids don't use it consistently unless someone is holding them accountable. The strongest results in the actual research come from combining the two, not picking a side.
That's a more useful answer than either side's marketing, and it's worth understanding why.
What Do People Actually Mean by "AI Tutor"?
This is where most comparisons go wrong before they even start. "AI tutor" gets used to describe at least three different products, and they don't perform the same way in the research.
The first is a conversational chatbot bolted onto an existing platform, like Khan Academy's Khanmigo, that answers questions and walks through problems on demand. The second is a standalone homework help app your child opens independently, with no teacher or parent checking in at all. The third is an AI-copilot that works alongside a live human teacher during a real class, flagging in real time where a student is stuck so the teacher can adjust, which behaves very differently from a chatbot a child talks to alone.
Adaptive practice platforms like IXL sit in a related but separate category. They personalize the sequence of problems using an algorithm, but they aren't conversational, and most parents use them as a supplement to instruction rather than a replacement for it. When a headline claims "AI tutoring works" or "AI tutoring doesn't work," it's worth checking which of these three it actually tested. The research below breaks down cleanly along that line.
What Does the Research Say About Human Tutoring?
Human tutoring has the deepest evidence base of any K-12 intervention. A 2024 meta-analysis in the Review of Educational Research, covering dozens of randomized controlled trials, found a pooled effect of about 0.29 standard deviations, roughly the difference between a student at the 50th percentile moving to the 61st. Effects were largest when a trained teacher or paraprofessional, not a peer or volunteer, delivered sessions at least three times a week, in earlier grades, during the school day.
Here's the part most "tutoring works" headlines skip. A 2026 analysis by researchers at Brown, Harvard, and Michigan, reviewing 265 tutoring studies, found that effect sizes shrink to roughly a third to a half of their original size once a program scales past 1,000 students in real public school settings, dropping from 0.42 standard deviations in the full research base to 0.14 to 0.21 under realistic large-scale conditions. Programs that kept the three-times-a-week, small-group, in-person structure held up better than those that didn't. In plain terms, tutoring isn't automatically effective just because it's human. How it's delivered matters almost as much as who's delivering it.
Here's the part most "tutoring works" headlines skip. A 2026 analysis by researchers at Brown, Harvard, and Michigan, reviewing 265 tutoring studies, found that effect sizes shrink to roughly a third to a half of their original size once a program scales past 1,000 students in real public school settings, dropping from 0.42 standard deviations in the full research base to 0.14 to 0.21 under realistic large-scale conditions. Programs that kept the three-times-a-week, small-group, in-person structure held up better than those that didn't. In plain terms, tutoring isn't automatically effective just because it's human. How it's delivered matters almost as much as who's delivering it.
What Does the Research Say About AI Tutoring?
The most-cited AI tutoring result comes from a 2025 Harvard study, later published in Scientific Reports, where undergraduates using an AI tutor for two physics topics learned more than twice as much in less time than students in a traditional class. It's a genuinely striking number, but the fine print matters: fewer than 200 college students, two narrow topics, and sessions that were scheduled and proctored over video call rather than used independently. It tells you AI tutoring can work under tight, supervised conditions. It doesn't tell you what happens when a 4th grader opens the same kind of tool alone after dinner.
That second question is exactly what a Stanford research team tested next, across two school districts serving grades 1 through 5. The result was blunt. Even with time set aside specifically for it, only 53 to 61 percent of students actually used the AI tutor, averaging just 2 to 5 minutes a week against a recommended 30. Pairing the tool with a human tutor who checked in on progress measurably increased how much students used it. Access to an AI tutor and actual use of one turned out to be two different problems.
A third study, run by the math platform Eedi with Google DeepMind, gets closest to a true head-to-head test. Students received either a human tutor, an AI tutor with human review, or generic hint text. Both human and AI-assisted support clearly beat generic hints, and AI-assisted tutoring performed about as well as human-only tutoring, with a small edge on helping students transfer what they'd learned to new problems. The detail that matters most for parents: human tutors still edited or rejected roughly a quarter of the AI's draft responses before they ever reached a student. The AI wasn't running the show. It was drafting, and a person was still checking its work.
That second question is exactly what a Stanford research team tested next, across two school districts serving grades 1 through 5. The result was blunt. Even with time set aside specifically for it, only 53 to 61 percent of students actually used the AI tutor, averaging just 2 to 5 minutes a week against a recommended 30. Pairing the tool with a human tutor who checked in on progress measurably increased how much students used it. Access to an AI tutor and actual use of one turned out to be two different problems.
A third study, run by the math platform Eedi with Google DeepMind, gets closest to a true head-to-head test. Students received either a human tutor, an AI tutor with human review, or generic hint text. Both human and AI-assisted support clearly beat generic hints, and AI-assisted tutoring performed about as well as human-only tutoring, with a small edge on helping students transfer what they'd learned to new problems. The detail that matters most for parents: human tutors still edited or rejected roughly a quarter of the AI's draft responses before they ever reached a student. The AI wasn't running the show. It was drafting, and a person was still checking its work.
AI Tutors vs. Human Tutors: Side by Side
| Factor | AI Tutor (chatbot or app) | Human Tutor (live) |
|---|---|---|
| Typical cost | Free to about $20/month | $25 to $80+ per hour |
| Availability | 24/7, instant | Scheduled sessions |
| Best-documented effect | Strong in short, proctored studies; unproven for independent home use | ~0.29 SD on average; up to 0.42 SD in ideal, small-scale programs |
| Biggest weakness | Low independent use without accountability | Effects shrink at scale; quality varies by tutor and program design |
| Judgment and relationship | Limited; can't reliably read frustration, motivation, or a bad day | Can adjust pacing, tone, and approach in real time |
| Best for | Practice reinforcement and quick homework questions for a motivated, organized student | Conceptual gaps, accountability, or a child who needs someone to notice when they're stuck |
Which should you choose? If your child is generally on track and needs occasional practice or a quick explanation, a free or low-cost AI tool is a reasonable first stop. If your child is falling behind, avoiding a subject, or needs someone to notice when they're stuck before they say so, the research points toward a live human tutor, and ideally one supported by AI tools rather than replaced by them. The strongest results in the research aren't AI instead of a person. They're AI reviewed by a person.
Signs the Tutoring You're Using Isn't Actually Working
The signs look almost identical whether the "tutor" in question is an app or a person.
• Your child opens the tool, clicks through a few screens, and closes it within 2 to 3 minutes most days
• They can get through practice problems with help but freeze on the same problem type on a worksheet or test without it
• Grades or homework completion are sliding even though the app shows completed lessons or a streak
• Your child says the tool "just gives the answer" rather than explaining the reasoning behind it
• Nobody, including you, is actually checking in on what the sessions are surfacing about where your child gets stuck
The signs that a human tutor isn't helping and the signs an AI tool isn't helping come down to the same underlying pattern: engagement without transfer. A session that produces a completed checkmark but no independent gain in the skill. The format doesn't change what to watch for.
• Your child opens the tool, clicks through a few screens, and closes it within 2 to 3 minutes most days
• They can get through practice problems with help but freeze on the same problem type on a worksheet or test without it
• Grades or homework completion are sliding even though the app shows completed lessons or a streak
• Your child says the tool "just gives the answer" rather than explaining the reasoning behind it
• Nobody, including you, is actually checking in on what the sessions are surfacing about where your child gets stuck
The signs that a human tutor isn't helping and the signs an AI tool isn't helping come down to the same underlying pattern: engagement without transfer. A session that produces a completed checkmark but no independent gain in the skill. The format doesn't change what to watch for.
What Should You Actually Do This Week?
• Sit with your child through one AI tutoring session and watch what they do when they get an answer wrong. Do they read the explanation, or click past it to the next hint?
• Time actual engagement, not subscription status. If your child spends under 10 minutes a week genuinely working with an AI tool, that's your answer, not the platform's usage claims.
• Ask your child to explain a concept back to you in their own words, no screen involved, right after a session. If they can't, the learning hasn't transferred yet.
• If you're using both an AI tool and a human tutor, ask the tutor directly whether and how they use AI-assisted resources, and whether they check the output before it reaches your child.
• Compare cost per actual hour of engagement, not the sticker price. A $20 a month app your child uses for 10 minutes a week costs more per useful minute than a single well-used tutoring session.
If you've tried monitoring and adjusting for three to four weeks and usage still hasn't improved, or grades still haven't moved, that's the point to bring in live, human support rather than switching to a different app.
• Time actual engagement, not subscription status. If your child spends under 10 minutes a week genuinely working with an AI tool, that's your answer, not the platform's usage claims.
• Ask your child to explain a concept back to you in their own words, no screen involved, right after a session. If they can't, the learning hasn't transferred yet.
• If you're using both an AI tool and a human tutor, ask the tutor directly whether and how they use AI-assisted resources, and whether they check the output before it reaches your child.
• Compare cost per actual hour of engagement, not the sticker price. A $20 a month app your child uses for 10 minutes a week costs more per useful minute than a single well-used tutoring session.
If you've tried monitoring and adjusting for three to four weeks and usage still hasn't improved, or grades still haven't moved, that's the point to bring in live, human support rather than switching to a different app.
Common Questions About AI Tutors vs. Human Tutors
Is AI tutoring as effective as a human tutor? The best current evidence suggests AI tutoring that's reviewed by a human performs about as well as human-only tutoring on math problem sets, according to a 2025 Eedi and Google DeepMind study. Fully independent AI tutoring shows much weaker real-world results, mainly because most kids don't use it consistently without some form of accountability.
What is the actual effect size of tutoring, AI or human? Human tutoring's average effect size across dozens of studies is about 0.29 standard deviations, though a 2026 analysis found that shrinks to 0.14 to 0.21 once programs scale to more than 1,000 students in typical public schools. AI-only tutoring hasn't been studied at that same scale yet.
Can an AI tutor replace a human tutor? Not based on current research. The strongest results come from AI tools that support a human tutor's work, not from AI tools used entirely on their own. A 2026 Stanford study found that even with scheduled access, most students used an AI tutoring app for only 2 to 5 minutes a week.
Are AI tutoring apps safe for kids to use unsupervised? Researchers and policy groups have raised real concerns about AI chatbots giving inaccurate or age-inappropriate responses without human oversight. Most current research on effective AI tutoring involves a human reviewing or approving the AI's output rather than a child using it fully unsupervised.
How much does a human tutor cost compared to an AI tutoring app? AI tutoring apps typically run free to about $20 a month. Live human tutors typically run $25 to $80 or more per hour depending on subject, credentials, and format.
What is the actual effect size of tutoring, AI or human? Human tutoring's average effect size across dozens of studies is about 0.29 standard deviations, though a 2026 analysis found that shrinks to 0.14 to 0.21 once programs scale to more than 1,000 students in typical public schools. AI-only tutoring hasn't been studied at that same scale yet.
Can an AI tutor replace a human tutor? Not based on current research. The strongest results come from AI tools that support a human tutor's work, not from AI tools used entirely on their own. A 2026 Stanford study found that even with scheduled access, most students used an AI tutoring app for only 2 to 5 minutes a week.
Are AI tutoring apps safe for kids to use unsupervised? Researchers and policy groups have raised real concerns about AI chatbots giving inaccurate or age-inappropriate responses without human oversight. Most current research on effective AI tutoring involves a human reviewing or approving the AI's output rather than a child using it fully unsupervised.
How much does a human tutor cost compared to an AI tutoring app? AI tutoring apps typically run free to about $20 a month. Live human tutors typically run $25 to $80 or more per hour depending on subject, credentials, and format.
How Cosmo Helps
Cosmo's model isn't AI instead of a live teacher, and it isn't a live teacher with no AI support either. Every class is taught live, 1:1, by one of Cosmo's teachers (fewer than 1% of applicants are accepted), while an AI-copilot works in the background during the session, surfacing exactly where a student is stuck so the teacher can adjust in real time. That's closer to the hybrid model the research above actually favors than either a pure chatbot or old-style tutoring alone.
It starts with a short diagnostic, 15 questions, that builds a Personalized Learning Plan from day one, so the first session isn't a guess. From there, progress gets checked through mini-assessments every four classes and a monthly report, not just a streak counter or a completed-lessons badge.
If you're trying to figure out whether your child needs more independent practice or the kind of live, responsive attention an app can't provide, a single diagnostic session usually makes that clearer than another few weeks of guessing. Try a free trial class today to see where your child actually stands.
It starts with a short diagnostic, 15 questions, that builds a Personalized Learning Plan from day one, so the first session isn't a guess. From there, progress gets checked through mini-assessments every four classes and a monthly report, not just a streak counter or a completed-lessons badge.
If you're trying to figure out whether your child needs more independent practice or the kind of live, responsive attention an app can't provide, a single diagnostic session usually makes that clearer than another few weeks of guessing. Try a free trial class today to see where your child actually stands.
New to Cosmo?
Your child's personalized learning experience starts here.
Try Cosmo