'Learn twice as fast' with AI? What Alpha School's numbers and real studies show

What's going on

Alpha School is a private school chain from Texas. Its pitch is simple: with two hours of software-based instruction in the morning, kids learn "twice as much as their peers." Afternoons are for workshops on topics like public speaking or teamwork. There are no traditional teachers, only "guides" whose job is to motivate. Tuition runs up to $75,000 a year. According to Fortune, supporters range from Silicon Valley to the Trump administration. Education Secretary Linda McMahon called the model "the most exciting thing I've seen in education in a long time."

Now, for the first time, internal numbers are out. The New York Times (Dana Goldstein, October 23, 2026) analyzed them, and Fortune picked them up today. At the same time, fresh controlled studies on AI tutors have landed. Our thesis: tightly guided AI tutors demonstrably help. An open-ended chatbot often doesn't, and can even do harm. And so far, the Alpha model is evidence for neither.

What Alpha's internal data shows

According to the NYT and Fortune reports:

The measurement ran from fall to winter, roughly half a school year. The available summaries don't say exactly how many children were included in the analysis. Alpha's own mid-year report from February 2026 lists about 407 "growth events" in math. We didn't find a statement from Alpha on the NYT figures in the sources available to us.

There's also criticism of day-to-day life at the schools. Former employees told 404 Media that the AI generated flawed lessons that sometimes did "more harm than good." Guides and parents told Wired that children worked late into the night to hit their daily goals. One mother described "stress and anxiety" in her 13-year-old to Fortune.

How the "2x" number is built, and why it tells you little

The basis is the MAP Growth test from the company NWEA. It's an adaptive test that adjusts to the child's level and measures performance in points on a continuous scale, called RIT points. According to critics, Alpha calculates it like this: the child's gain divided by the national median gain. If you gain 14 points when 7 is typical, you've hit "2x." Alpha's website doesn't disclose the formula, the sample or a comparison group.

This ratio has three weaknesses:

1. Small denominators inflate results. In higher grades and over short periods, the typical gain is small. Alpha's own report gives a 9th-grade example: a 9-point gain against a norm of 2 yields "4.5x." For 8th-grade math, NWEA's 2025 norms overview lists a median gain of 4 points between fall and winter, with a standard deviation of 8 points. So individual children vary twice as much as the average even grows. Our back-of-the-envelope math: if the distribution were roughly normal, about 30 percent of a completely ordinary student body would gain at least 8 points — i.e. hit "2x." Part of that spread is measurement noise. The calculation is an approximation, but it shows how shaky a "2x" is for an individual child.

2. There's no comparison group. Alpha compares against the national median, not against similarly affluent and motivated children. The Fordham Institute calls such numbers without a comparable group "nearly meaningless," especially since families paying this kind of tuition deliberately chose the school. Teacher and blogger Dylan Kane makes a similar argument: kids from families paying $40,000 or more probably learn faster than the median anyway.

3. There's a better measure, and it comes from the test provider itself. NWEA offers the "Conditional Growth Index": gain minus the expected gain for children with the same starting level, divided by the spread. This metric accounts for where a child started. To be fair, Alpha and authors close to it do report such values too, some very high, between 2.0 and 7.86. A commenter under Kane's post notes that kids at every achievement level grow two to four times faster. That would argue against a pure selection explanation. But all of these numbers are self-published. The school's leadership explicitly had Gemini and Claude analyze the mid-year report. According to Scientific American, Alpha has never released the raw data.

Our take: the 37 percent among weaker math students is especially telling. Students who start low often improve at the next measurement for purely statistical reasons — an effect called regression to the mean. That group in particular should have done well.

What happens when there's no selection

A counter-test already exists. Unbound Academy is a public online charter school in Arizona that uses Alpha's "2 Hour Learning" and has to accept every child. According to an analysis by math teacher Dan Meyer, it reached 10 percent math proficiency after its first year (projection: 60 percent, state average: 34) and 28 percent in reading (projection: 65, average: 42). Meyer concedes that the students may have started out weaker. His conclusion:

"It's far harder to make successful students than to select them in advance."

According to the Texas Tribune, only 21 percent passed state tests at an Austin charter school using the program. At least nine states rejected Alpha charter applications. In January 2025, Pennsylvania called the AI model "untested." The network keeps growing anyway: according to company documents, more than 2,600 public school students in Texas and South Carolina use the platform. Houston, Aldine and Fort Davis are piloting it as a supplement to instruction. In fall 2026, Drew University in New Jersey opened a K-8 school on its campus with Alpha co-founder Brian Holtz, with tuition starting at $30,000. Students and alumni are protesting.

One important point for the chatbot comparison: Alpha's software, TimeBack, is mostly adaptive practice software with a reward system. In the previous year, according to Kane, the math app IXL was the centerpiece — "practically no generative AI." He suspects the real "secret sauce" is motivation and incentives, not AI.

What controlled studies actually measure

Randomized controlled trials (RCTs) are more reliable. In these, chance decides who gets the tool. Effects are reported in standard deviations (SD), a measure of how far a group shifts relative to the spread of all results. In education research, 0.2 SD already counts as substantial.

The Maryland study is the largest. Instructors were randomly assigned to get the assistant or not. The −0.37 SD figure (about four grade points) comes from comparing identical courses with 1,353 students. In the full sample, the grade effect is smaller and only barely significant. The drop in course engagement, however, remains clear. Among first-generation college students, the losses were more than twice as large.

What stands out is how little it was used. Only about 15 percent tried the assistant. 73.8 percent of requests were for information, explanations or solutions; only 0.7 percent asked for feedback on their own attempt. Most instructors kept the default mode, which answers directly. So what was measured is the effect of offering the tool, not of intensive use. The authors suspect the assistant crowded out work with the existing course materials.

The pattern across all the studies is similar. Gains showed up where the tutor was tightly tied to the curriculum, assignments and flow, often with humans alongside. The World Bank authors explicitly say they can't attribute the effect to the chatbot alone. Where learners use an answer machine freely, the benefits vanish or reverse. And even good tutors only work if people use them: in Tennessee, the median student wrote anything to Khanmigo on just a third of practice days.

And in Germany?

In Germany, the state-run school chatbot telli — since renamed AIS.chat — has been rolled out across the federal states. According to news4teachers, the overarching AIS project has a budget of 50 to 60 million euros. We found no impact study on learning outcomes. The GEW teachers' union and the Saarland teachers' association see the tool playing only a "subordinate role" so far. Caution is also warranted with success stories: a widely cited meta-analysis that credited ChatGPT with a "large positive impact" on learning performance has been retracted, according to t3n. It had appeared in a journal from the Nature publishing group.

The bottom line

Alpha's "2x" is a marketing metric, not proof of impact. It divides short-term test gains by a national median, without a comparison group and in a heavily selected student body. Even by this generous calculation, most students miss the goal. Where Alpha's software was used without selection, results stayed below the state average. That doesn't prove the model is useless: small groups, lots of practice time and strong incentives can certainly work. But it isn't proven. Public school districts are adopting the software before anyone has independently evaluated it.

Our take: the research is more nuanced than "AI helps" or "AI hurts." There's evidence for tightly designed tutors that ask questions instead of giving answers, are tied to curriculum and assignments, and are embedded in instruction with humans. The effects are closer to 0.1 to 0.3 SD than to a doubling, measured over weeks rather than years. For schools, parents and education ministries alike, that means: access to a chatbot isn't what matters — how the tutor is designed and integrated is. And anyone advertising learning gains should open their data to independent researchers.