The Number They Won't Release

On Monday, July 13, Gerald K. LeTendre, a professor of educational administration at Penn State, published a short essay in The Conversation.¹ UPI and phys.org syndicated it the same day.² The subject was Alpha School, the for-profit AI academy that has grown in three years from a single Austin campus to more than fifteen locations across the country, with a Boston site opening this year and tuition running from roughly forty thousand to seventy-five thousand dollars a year.³

Alpha's central claim is famous inside the AI-in-schools world by now. Students spend two hours a day on AI-driven mastery work in reading, math, and other core subjects, and, the school says, learn more than twice as fast as peers in a conventional six-hour day. The evidence Alpha most often points to is its own NWEA MAP Growth data, which it describes as showing gains roughly 2.6 times larger than similarly scoring students elsewhere.³

LeTendre made one small observation. Nobody outside the company has seen the underlying data.¹

The Argument Underneath

The intellectual heritage Alpha invokes is Benjamin Bloom's 1984 "2 Sigma Problem," the finding that one-on-one mastery tutoring could move an average student roughly two standard deviations up the achievement curve. Alpha's pitch, roughly, is that a good AI tutor is the digital fulfillment of the mastery model Bloom described. The pitch is a real intellectual bet, not a marketing slogan, and it deserves a serious answer.

The serious answer, at the moment, is that the peer-reviewed literature does not support it. A large 2020 review by the National Bureau of Economic Research found consistent, meaningful gains from tutoring provided by human beings across ages, subjects, and settings.¹ A 2026 Brookings review of generative AI in tutoring found that the new tools do improve on the older computer-based tutors, mainly by letting students interact in ordinary language and by adapting mid-session.⁴ Neither review found evidence that AI tutors outperform skilled human tutors. What they found is that AI tutors can be cheaper, more available, and more patient at three in the morning. That is a meaningful contribution. It is not the same as being better.

LeTendre's essay does not say Alpha is lying. His argument is quieter and more useful. Claims of this size, in a field this important, need to be verifiable by people who do not stand to profit from them. Alpha's numbers might be right. Nobody outside Alpha can say.

What Actually Grew

The most interesting number in this week's story is not 2.6x. It is fifteen. Alpha went from one school to more than fifteen in three years, funded almost entirely by families who read the marketing and made the decision the way families always make school decisions, on some mix of hope, exhaustion, and a story that sounded like relief.³ That is the real growth curve. It is going to be repeated by dozens of other AI school models over the next several years, at every price point, with every variation of the same claim.

None of those schools will release their data either, unless a norm forms that says they have to. The absence of that norm is the actual education story of this decade. Public schools live and die by the numbers. State report cards, NAEP, MAP, ACT, everything is public, contested, sliced by subgroup, argued over in board meetings that run past midnight. The new AI schools, in the same market, at higher price points, with more experimental methods, are being asked for nothing of the kind. Alpha's press releases and its parents' testimonials are, for now, the record.

What A Score Can And Cannot Say

Even if Alpha did release the MAP data tomorrow, and even if the gains held up, the deeper question would not close. A test score is a snapshot. It cannot tell you how the student got there. Whether she wrote a sentence and then rewrote it. Whether she read a passage twice. Whether she paused before answering a question because she was thinking, or because she was waiting for the tool to think for her. Two children can arrive at the same score by very different paths, and the path is what the next teacher will need.

That is the layer we spend our days on at Koan. Not another AI tutor to add to the market, but the quiet substrate underneath the tools already in the room, the one that captures what a student actually did while she was working. The revisions kept and the ones thrown away. The moment she asked for help and the moment she chose not to. The half-drafts, the pauses, the small refusals to accept an answer she did not understand. Those artifacts are what turn a score into a story a teacher can teach into.

The Alpha argument, and the broader argument it stands in for, will not be settled by press releases. It will be settled by making the work visible enough that the settlement can happen in public. Right now, in most of the schools trying most of the new tools, the record of the child's thinking is not being kept at all.

If the fastest-growing school model in the country will not show its numbers, what makes us think the slower ones will?