The data is unambiguous: students using AI tutoring achieve test scores 54% higher than those in traditional classrooms. Khan Academy's three-year efficacy study in Newark, New Jersey tracked 5,200+ students and found that every 10 skills learned on the platform correlated with roughly one point gain on the state test. Scale that up—a student mastering 60 additional skills in a year could expect 6-8 point gains, triple the average state improvement of 2 points.
This should be the end of the debate. It isn't.
Because behind the headline numbers lies a more complex reality: AI tutoring works brilliantly at one thing (drilling skills, providing 24/7 access, generating personalized practice), but it's not a replacement for human instruction. When you combine human tutors with AI, the results jump again. And teachers across the country are quietly panicking that test score gains might be masking something deeper—that students drilling skills with AI could be losing the ability to think critically or solve novel problems.
The tutoring market is betting everything on scale. The global AI tutoring sector is projected to grow from $1.63 billion in 2024 to $7.99 billion by 2030—a 30.5% compound annual growth rate. Khan Academy alone grew from 68,000 users in the 2023-24 school year to 1.4 million by mid-2025, and expanded partnerships from 45 school districts to 380+. Adoption among students is nearly universal: 86% of students globally now use AI in their studies, with ChatGPT (66%) and Grammarly (25%) leading the pack.
The question isn't whether AI tutoring works. It does. The question is what it's actually optimizing for—and what we might be losing in the process.
The Effectiveness Story (It's Real)
Khan Academy's study is the gold standard here. Over three years, researchers tracked students in Newark public schools using Khanmigo (Khan Academy's AI tutor) alongside traditional instruction. The correlation was linear and measurable: more skills learned, more test score growth. A student who learned 60 additional skills could expect gains of 6-8 points on New Jersey's state assessment—a meaningful improvement in a system where average annual gains hover around 2 points.
This isn't an outlier. Rising Academies, which operates in East Africa, deployed an AI math tutor called Rori via WhatsApp (no internet required) and saw learning gains of 0.3 standard deviations—equivalent to about one year's worth of typical progress. Stanford's TutorCoPilot research found that AI-assisted human tutoring enabled weaker tutors to nearly match the effectiveness of expert tutors. The pattern is consistent: AI tutoring works.
The mechanism is straightforward. AI tutors provide instant feedback, adapt to individual pacing, and generate unlimited practice problems. A student struggling with fractions gets targeted drills. A student ready to advance moves on. No waiting for a teacher to get to you in a class of 30. No shame in asking the same question for the fifth time. The personalization is genuine.
Student perception backs this up. A Chegg survey from 2025 found that 50% of students reported improved understanding of complex concepts when using generative AI, up from 44% in 2023. Adoption isn't a curiosity anymore—it's mainstream. 84% of high school students now use generative AI for schoolwork.
But here's where the story gets interesting.
The Hybrid Reveal: Humans Still Matter
In September 2025, Carnegie Mellon University published research that should have gotten more attention than it did: students tutored by humans plus AI outperformed students tutored by AI alone. The gains increased with time on task. Human tutors enhanced the benefits of AI tutors.
This is not a small effect. This is a fundamental limit of pure AI tutoring.
The researchers' interpretation: AI tutors can't fully replicate the emotional support and adaptive dialogue that human tutors provide. An AI can generate a practice problem and check your answer. A human tutor can sense frustration, adjust their tone, tell you a story that makes the concept click, or recognize that you need a break. That's not a feature that can be optimized away.
This creates a market opportunity that the venture-backed pure-play AI tutoring platforms haven't fully reckoned with: the hybrid model. Not "AI replaces tutors" but "AI amplifies tutors." A human tutor with AI assistance can handle more students, generate better explanations, and spend less time on grunt work. As we covered in Legal AI's ROI Problem, the real value isn't in automation—it's in augmentation. The same principle applies here.
The Hidden Cost: What Happens to Thinking?
But teachers aren't celebrating. They're worried.
According to Michigan Virtual's 2025 snapshot of AI in education, educators are expressing "deep worries" that AI tutoring could diminish students' critical thinking and problem-solving abilities. The concern isn't theoretical. If a student can ask an AI tutor to explain any concept instantly, do they ever struggle with the productive confusion that builds understanding? If they drill problems until they pass the test, do they learn to think about why a solution works?
The worry goes deeper. PhD students surveyed in the research expressed concerns that AI trained on historical content could embed biases into responses, reinforcing stereotypes and conventional thinking. One observation stood out: "Loss of Innovation" from over-reliance on AI. If students outsource thinking to a tool, what happens to creativity?
This is the paradox: test scores up, thinking potentially down. It's the same dynamic we see in other AI-augmented fields. The tool optimizes for measurable outcomes (test scores, speed, efficiency) while potentially degrading unmeasured ones (creativity, resilience, deep understanding).
The research doesn't prove this happens—it's a concern, not yet a demonstrated harm. But it's a concern from people who work with students every day. And it's worth taking seriously.
The Adoption Reality Check
Teachers are adopting AI tools, but not seamlessly. 60% of K-12 teachers have adopted AI tools as of 2025, but Michigan Virtual's research identified major barriers: limited training, time constraints, unclear district policies, and concerns about inappropriate student use and overdependence on technology.
In other words: the tools exist, adoption is happening, but integration is messy. Teachers are using AI, but they're not yet confident they're using it right.
Privacy is another friction point. Parents and educators worry about student data in AI systems. Most major platforms (Khan Academy, Chegg Tutors, Photomath) collect learning data to improve recommendations—which is valuable for personalization but raises legitimate questions about who owns that data and how it's used.
The Market Opportunity (And the Trap)
The numbers are seductive. A $1.63 billion market growing at 30%+ annually is venture-capital catnip. Every major edtech company has launched or acquired an AI tutor: Google has Socratic, Chegg has integrated AI across its platform, Khan Academy has Khanmigo, Quizlet has AI study tools.
But the Carnegie Mellon research suggests the real opportunity isn't pure AI tutoring—it's infrastructure for hybrid models. Tools that make human tutors more effective. Platforms that combine AI skill-drilling with human mentorship. Systems that use AI to handle the commodity parts (generating practice problems, checking work, explaining concepts) so humans can focus on the non-commodity parts (motivation, debugging confusion, building resilience).
That's a harder business model to scale. It can't be pure software. It requires coordination with human tutors, schools, and districts. But it might be the model that actually works.
What This Means
The AI tutoring market is real and growing fast. Test score gains are measurable. Student adoption is nearly universal. But the story isn't "AI replaces tutors." It's "AI changes what tutoring looks like, and we're still figuring out what that means."
The companies betting on pure AI tutoring are capturing market share and user growth. But the ones that will own the category long-term are probably the ones that figure out how to make human tutors superhuman, not replace them. That's a slower story, but it's the one the data actually supports.
For students: AI tutoring works for what it's designed for—skill-building and practice. But it's not a substitute for human guidance, especially when the goal is deeper learning.
For educators: the tool is here. The question now is integration. How do you use AI to handle the repetitive parts without letting it atrophy the thinking parts? The answer probably isn't "ban it" or "use it everywhere"—it's "use it strategically, with human oversight, as part of a broader learning strategy."
For investors: the market is real, but the moat is human relationships, not algorithms. The winners will be the platforms that make schools and tutors more effective, not the ones that try to replace them.
Originally published on Derivinate News. Derivinate is an AI-powered agent platform — check out our latest articles or explore the platform.
Top comments (0)