A year ago, I wrote a rant about Alpha schools. A few times a year, I let myself go. No, this will not be a new rant. Sorry. But in my post and the follow-up post, I argued that we couldn’t separate Alpha School’s spectacular claims from the extraordinarily selective population it serves. Now we seem to have some possible evidence. Does that mean we are finally getting to the alpha and omega of Alpha School? Probably not. But we do seem to know a little bit more.

The main reason I am returning to the Alpha School story is a new article in Scientific American that asks a fairly straightforward question: does its model actually work?

If you have missed my previous posts about Alpha or haven’t read any other sources yet, the basic idea sounds almost too good to be true. At Alpha Schools, students supposedly learn the core academic subjects such as language and maths in a mere two hours a day. They do this largely through adaptive software and AI. Teachers are replaced by ‘guides’, whose role is less about teaching content and more about motivating students and helping them stay on track. Forget about coaches; guides are the new trend. The rest of the school day is than filled with entrepreneurship, public speaking, and other life skills, like gaining followers on Twitter/X.

And Alpha has been claiming impressive results. Students supposedly grow 2.6 times faster than their peers on nationally normed MAP tests.

I don’t want to dismiss the underlying idea by definition, since this isn’t a rant post. Scientific American also reports on research suggesting that intelligent tutoring systems and adaptive learning software can improve learning. The article mentions, for example, a recent Harvard experiment that found impressive learning gains with a carefully designed AI tutor. The researchers are also mentioned in the article. They seem to be very cautious themselves.

Because… there is a giant leap between saying that adaptive tutoring can help students learn and saying that you can replace much of regular classroom instruction with two hours of software. A small step for mankind, that kind of leap. And this is where the Scientific American article becomes much more cautious.

The impressive results Alpha presents have not been independently verified in a way that allows us to conclude that the model caused them. One needs to know that Alpha’s students are also a very particular group. Tuition at many of its schools ranges from $40,000 to $75,000 a year, as mentioned in my first rant. If you can afford such a school, it’s unlikely that you’re part of random sample of parents, and neither are your children.

So the most important conclusion from the article seems to be that we still don’t really know whether Alpha School works.

A bit of a boring post so far. But wait, do we really know so little? I don’t think so!

Shortly before the Scientific American article appeared, Dan Meyer published a fascinating post asking a slightly different but connected question: Does Alpha School work for regular kids? And he found something that comes surprisingly close to a natural test of the criticism I made a year ago.

In his post, Meyer first documents just how selective Alpha School actually is. It’s more than money. Prospective students participate in a so-called ‘Shadow Day’. Other elements that you need to know:

• Older students may have to submit additional work.
• Parents need to be sufficiently engaged to choose and navigate an unconventional school model (reminds me of some modern schools I know)
• Students need to be willing to attend (of course, but still not common with every school/every student)

Alpha’s application also states that students should function within a typical range of ability and independence. Applicants should not have any serious prior disciplinary issues. Maybe logical too, but how many schools wish they could ask this too? Alpha looks for students who can give and receive feedback, contribute to the school’s culture and, importantly, be ‘coachable’.  None of this means that Alpha’s students don’t learn a lot. But it makes the causal question about their approach rather difficult. Does Alpha make unusually successful students? This all seems to suggest that Alpha is unusually successful at selecting students who are likely to succeed.

Cue Unbound Academy. Unbound Academy is a public charter school in Arizona that opened last year, in 2025. In its charter application, the school explicitly connected itself to Alpha School and adopted the same basic 2 Hour Learning model with AI-driven adaptive learning technology, condensed academic learning in the morning, life-skills workshops in the afternoon, and teachers reconceived as guides. We can put a check next to this. But there is one important difference.

As a public charter school, Unbound cannot select its students in the same way Alpha can. This means that parents don’t need to pay $40,000 or $75,000. Also students don’t have to meet Alpha’s criteria concerning independence, behaviour or coachability.

No, this isn’t a randomised controlled trial, those parents still chose this kind of school, which also can have an effect. But it does give us something we didn’t have before: something resembling a quasi-experiment in which much of the Alpha model remains while a major part of the selection mechanism disappears! So does Unboud Academy get the same results?

In its application, Unbound Academy had clear goals. It predicted that 60 per cent of its students would reach proficiency in mathematics in their first year, compared with a baseline expectation of 34 per cent. For English Language Arts, the target was 65 per cent, compared with a baseline of 40 per cent. The longer-term ambition was even more striking: proficiency rates above 90 per cent within two or three years.

Do note, those predictions weren’t imposed on the school by skeptical outsiders like, ehm, me. They were the school’s own targets! Everything is being justified in part by the results achieved at affiliated schools using the model.

But a year later, mathematics proficiency isn’t 60 per cent; it is 10 per cent. English Language Arts proficiency? No, not 65 per cent, but a mere 28 per cent. If we look at these results, they are also below the Arizona state averages of 34% and 42%, respectively. Not really close. So no cigar.

But wait, before anybody declares the great AI-school experiment dead, there are several reasons to be careful. Yes, I’m getting all nuanced again. One should start wishing for a rant post. We need to remember that this is one (pretty new) school. It is its first year. We don’t know whether the students who actually enrolled started at the baseline levels assumed in the charter application. And none of this demonstrates that students would have learned more in a conventional school. We just don’t have enough information.  So these results do not show that the Alpha model makes children learn less. Hell, this data doesn’t even prove that the model cannot work in a public school.

Dan Meyer is also careful about this. As he puts it, these results shouldn’t be the last word on 2 Hour Learning, but he and I are convinced that they probably should be taken much more seriously than they have been so far.