What the Leaked Documents Actually Show

Leaked internal documents from Alpha School, recently reported by Business Insider, reveal something more structurally interesting than the headline suggests. The AI tutoring system at the center of Alpha's model is generating lessons that school administrators themselves describe as doing "more harm than good" in some cases. The Trump administration has publicly praised Alpha School as a model for AI-integrated private education. The leaked documents complicate that endorsement considerably.

What strikes me about this story is not that the AI is producing bad outputs. That is a known and tractable problem. What strikes me is the institutional framing around it. Students are absorbing faulty instructional content from a system positioned as authoritative, and the governance architecture for catching and correcting that content appears to be underdeveloped. This is not primarily a machine learning problem. It is an organizational theory problem.

The Awareness-Capability Gap in Institutional Form

My dissertation research focuses on what I call the awareness-capability gap: the well-documented finding that knowing an algorithm exists, and even knowing something about how it works, does not translate into the ability to respond to it effectively (Kellogg, Valentine, & Christin, 2020). Most algorithmic literacy research treats this gap as an individual cognitive failure. The Alpha School case suggests the gap operates at the institutional level as well.

Alpha School's administrators are presumably aware that their AI system can produce errors. The leaked documents suggest they have communicated this awareness internally. But awareness did not produce corrective capability. The governance infrastructure - the feedback loops, the human review protocols, the escalation procedures for flagging faulty content - does not appear to have kept pace with deployment. Awareness and capability are not the same thing, whether the unit of analysis is a gig worker or an educational institution (Gagrain, Naab, & Grub, 2024).

Routine Expertise and the Substitution Error

There is a deeper theoretical issue here that connects to Hatano and Inagaki's (1986) distinction between routine and adaptive expertise. Routine expertise is the ability to execute well-defined procedures reliably. Adaptive expertise is the ability to recognize when procedures are failing and to construct novel responses. The pedagogical model at Alpha School, at least as described in the leaked documents, appears to assume that AI tutoring tools require routine expertise to operate: teachers or supervisors following a defined process of content delivery. What the faulty lesson problem reveals is that the environment actually demands adaptive expertise - the capacity to evaluate AI outputs against first principles and intervene when those outputs are structurally wrong.

This is the substitution error I see repeatedly in AI deployment narratives. Organizations treat AI systems as procedure-replacers when they should be treating them as procedure-generators that still require principled human evaluation. The distinction matters enormously in high-stakes domains like education, where Sundar (2020) has shown that machine-generated content carries an implicit authority signal that can suppress critical evaluation by recipients, including students who have been told the system is designed for them.

What Governance Actually Requires Here

The Tailscale announcement this week about identity-linked governance for AI agents points in a useful direction, even though it is aimed at enterprise security rather than education. The core insight is that AI governance requires auditability at the output level, not just at the access level. Knowing who used the system is not the same as knowing what the system produced and whether that output was valid.

Applied to Alpha School's situation: the governance gap is not about whether administrators can see that AI lessons were delivered. It is about whether anyone with domain competence is reviewing the content of those lessons against pedagogical standards before students receive them. That is a workflow design problem, and it is one that organizational theory has useful things to say about. Hancock, Naaman, and Levy (2020) argue that AI-mediated communication environments shift accountability in ways that are not intuitively obvious to participants. In this case, accountability for instructional quality has been partially delegated to a system that cannot hold it.

The Broader Signal

The Alpha School case is worth watching not because AI tutoring is inherently problematic, but because it is an unusually well-documented example of what happens when deployment velocity exceeds governance design. The political attention the school has received - from both press and administration - adds a layer of institutional pressure that makes honest internal evaluation harder, not easier. Organizations under external validation pressure have documented tendencies to suppress negative internal signals (Polychroniou, Trivellas, & Baxevanis, 2016). If that dynamic is operating here, the students described in the headline as guinea pigs are not incidental casualties of a technology experiment. They are participants in an institutional failure with a recognizable theoretical structure.