When Carl Hendrick took his first ride in a Waymo, the self-driving taxi service, he thought it felt a lot like the future of education.
Hendrick is a senior learning scientist at Alpha School, a buzzy private-school company where he trains artificial intelligence models to teach students—a process, he says, that is not unlike training a self-driving car.
For such a car, “it was just thousands and thousands of miles, training itself on data, and not just on a linear path. The challenge is in all the different eventualities, like the cyclist who’s wobbling in the road, the kid running out,” he says. “I think, in five or 10 years, a similar thing will happen in education where we train AI models on learning and all of kids’ misconceptions and problems.”
On supporting science journalism
If you’re enjoying this article, consider supporting our award-winning journalism by subscribing. By purchasing a subscription you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today.
For Alpha School, the equivalents to the swerving bikes and unexpected pedestrians are more like misconceptions, learning differences or preteen classroom struggles. And the “cars” the company is testing have 10-year-olds in the backseat.
In theory, life at Alpha School looks like this: students spend two hours each morning—or more, if they don’t meet their benchmarks—working with an AI tutor that adapts to each student’s needs, assesses their progress and gives live feedback. Hendrick and the company’s other learning scientists work with software developers to train the models using the latest research and curriculum tools. In the afternoon, students work on “life skills” by attending workshops on topics that include coding, entrepreneurship and public speaking. Tuition at most campuses ranges from $40,000 to $75,000 per year.
And now Alpha is expanding. This fall the company plans to grow to roughly 50 campuses across the U.S., including 27 newly announced locations.
Joe Liemandt, Alpha School’s billionaire principal, wants to “reach a billion kids,” Hendrick says. Hendrick himself believes the approach could help solve problems facing educational systems today.
One in five U.S. students is chronically absent, and reading and math scores have been declining across the country for a decade. When used by a teacher, adaptive learning models like Alpha School’s have shown promise in improving student performance and understanding. Alpha takes the idea considerably further, handing over much of academic instruction to software.
Alpha School has not released the underlying data behind its claims. But Dan Goldhaber, an education researcher at the University of Washington, who saw the school’s math instruction model last year, says it resembles other, established adaptive-learning technologies such as IXL Learning. Those technologies and other AI tutoring models have been studied publicly.
In a 2025 randomized experiment conducted in an introductory physics course at Harvard University, some students used an adaptive AI tutoring model that the professors had trained, while others used in-class active learning strategies. When students used the tutoring model, the researchers found, their median learning gains were more than twice as high.
Kelly Miller, a Harvard professor, who co-authored the study, says Alpha School’s approach, which centers on active learning, is “better than what currently exists in K–12.”
“The future of education is adaptive learning,” she says.
But Miller uses AI and recorded lectures to cover basic material so that her in-person time with students can be spent on more complex work. If there’s no teacher-student relationship—like at Alpha School, where adults supervising students generally aren’t trained educators—“what’s the point?” she says.
Greg Kestin, lead author of the Harvard study, has similar concerns. If students learn and are assessed under similar conditions, he says, they can develop “brittle knowledge,” performing well in that setting without being able to apply what they’ve learned elsewhere.
Gerald LeTendre, an educational policy researcher at Pennsylvania State University, who did not work on the study, says the “basic premise” of Alpha School is not wrong. But good tutoring and good teaching, he emphasizes, are not equivalent. “There’s no evidence in the literature that you can substitute a pure AI tutoring model and still achieve all of the multiple effects that one gets from studying in a classroom.”
A 2025 literature review of 28 studies that involved a total of nearly 5,000 K–12 students found generally positive effects from AI-driven intelligent tutoring systems—though the advantage was negligible when compared with nonintelligent tutoring systems that also used active learning models without relying on AI.
Hendrick admits Alpha School’s model is “not perfect.” But he argues that most kids aren’t getting a good teacher or instruction grounded in learning science in a classroom anyway. “You have massive variance in teacher quality and pupil outcomes in a way that most people would find shocking,” he says. “The variance within schools between teacher quality is often greater than the variance between schools. If you apply that to doctors or dentists, there’d be a public outcry.”
That’s Hendrick’s biggest argument for Alpha School’s model as a solution: “You can’t scale good teaching,” he says. Software, Hendrick argues, can be updated across classrooms at once as new curriculum and instructional research emerges and can give feedback far faster than a teacher grading assignments one by one.
Software, of course, scales quite nicely—and so do its mistakes. This year a 404 Media investigation reported that some of the lessons in Alpha School’s model were faulty; one former employee said, “Students were being treated like guinea pigs.” Alpha spokesperson Anna Davlantes says Hendrick’s team spends its days “poking holes in everything constantly to try to see: How do we get better?”
If you ask Alpha School, it’s working. The data it has released, measurements of students’ math and reading skills based on the Northwest Evaluation Association’s (NWEA’s) Measures of Academic Progress (MAP) framework, show strong results. But that may reflect “who’s attending the school, not the experience of the kid attending,” says Andrew McEachin, an educational policy researcher at the Educational Testing Service, who helped develop NWEA’s MAP system.
“These reports aren’t designed so that you can draw specific inferences from them,” McEachin says. “You can’t point to a specific policy or program or school system and say these scores are at this level because of this reason.”
Miller wants to see a “more rigorous analysis” of the process before drawing conclusions about what makes Alpha work. “It’s not that I don’t believe that they’re successful,” she says. “But I think to scale it, they need to understand a bit better why it’s successful and what are the things that are crucial in making it successful.”
That question is becoming more pressing as Alpha grows. The company is also beginning to test versions of its model in public schools—two in Houston and one in Springfield, Mass.—where it will encounter a much broader mix of students. Those new settings will put Hendrick’s premise to the test: If good teaching can’t scale, can good instruction?
