AI Tutoring with Khanmigo in a Two-Year School Experiment

TEXT START: Generative AI has been promoted as the technology that could transform education by providing every student a personal tutor.

The Dissection

This is not evidence of AI replacing teachers, and it is not strong evidence that AI adds value beyond ordinary educational software. The trial inserted Khanmigo into existing remedial mathematics sessions. Students mostly ignored it; when they used it, their interactions were usually shallow. The reported gains resembled Khan Academy practice without AI assistance.

The text’s central maneuver is to convert weak AI-specific results into a conditional success narrative: the technology is presumed promising, while failure is assigned to student engagement. The machine is not tutoring at scale. It is waiting for students to activate it.

The Core Fallacy

The text treats engagement as an external implementation problem rather than part of the product’s economic mechanism. A tutor that requires sustained, substantive dialogue from students already struggling in remedial sessions has not solved tutoring. It has moved the activation problem onto the least activated users.

“Give students access, then make them engage” is not a minor deployment detail. It is the product. The study does not establish that greater use would produce proportionally greater gains, nor that AI tutoring is superior to non-AI practice.

It also confuses educational improvement with productive participation. Even a 0.14-standard-deviation gain would be a narrow learning effect, not a restoration of the wage-to-consumption circuit. It does nothing to establish Sovereign ownership, Servitor indispensability, or resistance to cognitive automation. P1 is not demonstrated; the result is ordinary software effectiveness wrapped in AI branding.

Hidden Assumptions

• More messages would generate proportionally larger learning gains.
• Students can be induced to engage at scale without intensive teacher supervision, incentives, or coercion.
• Test-score gains are a reliable proxy for durable capability and future economic necessity.
• The main limitation is student usage rather than tutor quality, interface design, incentives, or task mismatch.
• Existing schools can absorb AI deployment without preserving the human labor structure that currently makes the deployment possible.
• Access will eventually become participation, and participation will eventually become productive economic value.

None of these assumptions is established by the supplied evidence.

Social Function

Partial truth functioning as techno-optimist anesthetic and transition management.

The text honestly exposes the engagement bottleneck, but uses that bottleneck to preserve the promise of AI tutoring. The implied message is that the technology works and only implementation remains. That is institutionally convenient: schools can treat AI as a supplementary experiment, vendors can classify weak results as an adoption failure, and nobody has to confront the harder possibility that the system has not produced a durable AI advantage at all.

The Verdict

This is an autopsy of current AI tutoring impotence, not a proof of educational transformation. Students were given an AI coach and mostly did not converse with it; outcomes tracked conventional Khan Academy practice. The machine has not become a tutor at scale. It remains an optional interface attached to human-managed schooling.

Under the Discontinuity Thesis, this is a lag instrument and a product-integration problem, not a new sovereign production system. If engagement can be engineered, learning gains may improve. That would still be a narrow capability gain, not a reversal of P1–P3 or a restoration of mass productive participation. The report’s hardest fact is also its most damaging: access is cheap, sustained use is scarce, and the study provides no evidence that AI knows how to command it.