AI Tutoring Went From Homework Shortcut to Institutional Tool, and Schools Are Learning the Difference the Hard Way

The education sector's first institutional response to capable conversational AI was defensive: detection tools, honor-code revisions, and in some districts outright bans, all premised on the technology being primarily a machine for outsourcing homework. The second phase now underway inverts the posture. School systems, universities, and education ministries in several countries are piloting or deploying AI tutoring systems deliberately built on the same underlying capabilities, on the theory that one-on-one tutoring — long documented as one of the most effective interventions in education, and long unaffordable to provide universally — might finally be deliverable at scale. The design distinction that separates the tools schools are adopting from the general-purpose assistants they initially banned is pedagogical rather than technical: tutoring configurations are constrained to guide rather than answer, withholding solutions in favor of hints, worked steps, and questions that probe where the student's understanding actually breaks. Early controlled studies of such systems have reported meaningful learning gains over unassisted study — with the recurring caveat that unrestricted assistants, the kind that simply produce the answer, show the opposite pattern: improved homework scores followed by weaker exam performance, the signature of practice outsourced rather than learning assisted. The implementation difficulties reported by early-adopting institutions cluster in predictable places. Students quickly learn to distinguish a constrained tutor from the unconstrained assistant in another browser tab, which makes the tutor's value proposition — genuinely better explanations, adapted to the student's demonstrated gaps — the real gatekeeper of usage. Teachers report the tools most useful for the middle of the distribution and least reliable for students with foundational gaps the system misdiagnoses, or for advanced students it fails to challenge. And the systems' occasional confident errors land differently in education than elsewhere: a student, by definition, is the user least equipped to notice. The realistic near-term settlement, visible in the programs that have matured past their pilot phase, positions the technology as an amplifier of instructional capacity rather than a substitute for it: unlimited-patience practice and explanation between human touchpoints, with teachers redirected toward the diagnosis, motivation, and relationship work that the systems demonstrably cannot do. That is a smaller claim than the transformation rhetoric that opened the decade, and a larger one than the cheating panic that followed it — which has been roughly the trajectory of every consequential educational technology before it.