Loading today’s stories
Science1 distinct publisher3 min readPublished
Alpha hands the academic core to software for two hours a morning at $40,000 to $75,000 a year, while the closest controlled evidence for the approach was gathered on undergraduates in a physics course.
The Scientist · Science desk
Compiled by The ScientistSomething wrong?How this is made
Carl Hendrick's own description of how a self-driving system gets good is thousands of miles of data covering the awkward cases, the cyclist wobbling in the road, the kid running out [1][2]. That is an argument about a measured record. The educational version of such a record would be student outcomes across many campuses and years, and the company has not released the data behind its claims [11]. Joe Liemandt's stated goal of reaching a billion kids, as relayed by Hendrick, is a scaling ambition rather than a finding [8].
Two design choices would make Alpha's internal numbers hard to read even if they appeared. Students who fall short of benchmarks get more tutor time, so instructional dose is assigned by performance [3], and any gain the school reports blends the software with the extra hours. The academic core sits inside that morning block, with afternoons given to workshops on coding, entrepreneurship and public speaking [4]. Price does the other filtering, at $40,000 to $75,000 in tuition at most campuses [5]. The national picture Scientific American sets beside the pitch, one in five students chronically absent and a decade of declining reading and math scores [9], is the comparison Hendrick's optimism invites [20] and the wrong denominator for a fee-paying, self-selected group.
The nearest thing to a controlled test comes from a course Alpha had nothing to do with. In a 2025 randomized experiment in introductory physics at Harvard, students assigned to an adaptive tutor the professors themselves had trained showed median learning gains more than twice those of students doing in-class active learning [13]. The control is the interesting part, because the comparison was against a well-run classroom rather than a lecture. What it does not tell you is whether the effect survives in ten-year-olds, across a full year, in subjects where the right answer is arguable. Greg Kestin, the study's lead author, adds the caution that bears hardest on a school where learning and testing happen in the same software: students assessed under the conditions they learned in can end up with brittle knowledge that does not travel [16].
Kelly Miller, who co-authored that trial, calls Alpha's active-learning emphasis better than what currently exists in K-12 [14]. She also describes using AI and recorded lectures to clear basic material so her in-person hours go to harder work, and asks what the point is when there is no teacher-student relationship, as at Alpha, where supervising adults are generally not trained educators [15]. That teacher-plus-tool arrangement is what the published literature actually covers: adaptive models like Alpha's have shown promise when a teacher is using them, and Alpha takes the idea further by handing much of instruction to software [10]. Dan Goldhaber of the University of Washington, who saw the math model last year, says it resembles established adaptive products such as IXL Learning, which have been studied publicly [12], so prior evidence exists and it describes a different staffing arrangement. Gerald LeTendre of Penn State puts the gap plainly: the basic premise is not wrong, but good tutoring and good teaching are not the same, and the literature does not show that a pure AI tutoring model substitutes for the rest [17].
Timing compounds the problem. Twenty-seven of roughly 50 campuses are newly announced, leaving about 23 older sites, so a bit over half the fall network will be in its first year [7], and first-year sites cannot produce cohort histories. My view, with its conditions: adaptive tutoring on well-specified procedural material, used by a teacher, has support worth acting on, while handing the instructional core to software for children as young as 10 [19] is a proposition that will be tested on paying families years before anyone outside the company can check it.
Ranked by verification strength, evidence, and original report placement.
Carl Hendrick is a senior learning scientist at Alpha School, a private-school company, where he trains artificial intelligence models to teach students, a process he says is not unlike training a self-driving car.
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.
In a 2025 randomized experiment in an introductory physics course at Harvard University, some students used an adaptive AI tutoring model the professors had trained while others used in-class active learning strategies; students using the tutoring model had median learning gains more than twice as high.
Gerald LeTendre, an educational policy researcher at Pennsylvania State University who did not work on the Harvard study, says the 'basic premise' of Alpha School is not wrong but that good tutoring and good teaching are not equivalent, and that there is no evidence in the literature that a pure AI tutoring model can be substituted and still achieve all of the multiple effects of teaching.
Scientific American characterises Alpha's deployment by writing that the 'cars' the company is testing have 10-year-olds in the backseat.
Alpha School students spend two hours each morning, or more if they do not meet their benchmarks, working with an AI tutor that adapts to each student, assesses progress and gives live feedback.
Distinct publishers with included, body-backed reporting in this cluster.
1 article · August 28, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
invest
A Connecticut judge just priced prompt injection: no fine, no e-filing2 distinct publishers
leadership
Robotaxis Are Taking Mid-Teens Share in Three Metros. Headcount Will Not Show It.1 distinct publisher
product
Uber's first European robotaxi still has a driver in it, and that is the whole story1 distinct publisher
science
Psilocybin's 'oneness' now has a brain-network correlate, and it was found by sorting people on self-report1 distinct publisher
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Strong reporting, thin proof
The claim under test — that software can carry the academic core for children — has no measurement attached to it anywhere in this reporting. Alpha has released no data. The one randomized trial cited was run on Harvard undergraduates in intro physics, by professors who trained the tutor themselves, and the 28-study K-12 review says the AI part adds almost nothing over active-learning tutoring without AI. What lifts the score off the floor is that the outside expertise is real and named: Goldhaber has seen the math model firsthand and places it alongside IXL, tools that have been studied in public.
Real classrooms, small footprint
This is past the demo stage. Two hours of software-led instruction every school morning, benchmarks that extend the session, and a fall plan for roughly 50 campuses are commitments with children and payrolls attached, at $40,000 to $75,000 a seat. But it remains a private-school network of unstated enrolment, roughly half of it opening cold, and the only signal from inside the deployment about quality is 404 Media's faulty-lesson finding reaching us secondhand. Liemandt's billion children are not adoption; they are intent.
A billion children, zero published results
The gap sits between two sentences of this story. One says the principal wants to reach a billion kids; the other says the school has not released its data. Around them, "the future of education is adaptive learning" arrives from a co-author of the trial that supports it, while the researcher with no stake in that trial says the literature contains no evidence you can swap in a pure AI tutor and keep what a classroom does. Scientific American is not the source of the inflation — it supplies the deflating quotes — but the marketing has run well ahead of anything measured on a ten-year-old.
Nearly every enthusiast has a stake
Trace who says the optimistic things. The narrator is on Alpha's payroll and trains the models he is describing. The rebuttal to the faulty-lesson report comes from Alpha's spokesperson. The professor who calls the approach better than current K-12 co-authored the study of a tutor her own team built and trained. Behind all of it is a billionaire owner and tuition of up to $75,000 a child. The genuinely disinterested voices — LeTendre, Goldhaber, and Kestin turning on his own study's design — are the ones raising doubts, which is the cleanest signal in this story about which way the pressure runs.
One outlet, but not one voice
Everything here comes through a single Scientific American piece, which caps how far we can go. It holds up better than most single-source stories because the reporting does the adversarial work itself — four named researchers from three institutions, a literature review that cuts against the pitch, and a second outlet's investigation acknowledged rather than ignored. The soft spots are the unattributed national statistics, the enrolment figures nobody supplies, and a company response that gets one line at the end.