AI in schools: who decides when it is ready?

AI in education briefing · 21 September 2026 · Developments reported 14–20 September

A school can make an AI tool available long before teachers, students and families agree about its place in learning. Two reports from the United States this week show how differently that negotiation can unfold.

In Midland, Michigan, a 15 September report described the implementation of an earlier SchoolAI adoption decision. Teachers in grades 6–12 choose whether and how to use the platform, with monitoring dashboards, activity boundaries and teacher-led training. This is an account of implementation, not evidence that learning has improved.

In Norwalk, Connecticut, more than 80 parents asked for a pause in elementary and middle-school AI use, an independent review and policy development free from commercial interests. Reported on 18 September, their intervention was a demand for change, not an enacted ban.

These cases do not establish a national trend. They do bring a practical question into focus: what would make a school community willing to try a tool, and what would justify continuing to use it?

For school-facing developers and consultants, the implication is that a pilot needs a clear educational purpose and an understandable account of who can intervene. Teachers need room to decline or adapt its use. Families need to know what happens to students’ data and what alternatives are available. The ability to pause should be a practical arrangement, with someone responsible for making the decision.

A useful observation is not yet a dependable judgement

A Hong Kong preprint posted on 16 September offers a more specific reason to keep evaluation close to classroom practice. Researchers compared GPT-5’s transcript-based ratings with human assessments of 87 observations from 38 classrooms in 30 kindergartens. Agreement varied across the observation framework, with greater divergence in procedural and context-dependent interactions.

The authors suggest using AI-assisted observation for preliminary screening and teacher reflection rather than high-stakes evaluation. The limits matter: this is a preprint, it concerns early childhood classrooms, and it measures agreement between ratings rather than gains in learning. It cannot establish that the same approach would work in secondary or IB settings.

For research into how students experience uncertainty with AI, that distinction is especially useful. A transcript may show a student asking for help; it does not, by itself, establish what that student understood or felt. An AI interpretation should remain something researchers examine alongside students’ own accounts. Disagreement between human and machine readings may reveal a question worth pursuing.

A promise worth testing

Microsoft’s draft Humanist AI Code of Conduct, published on 14 September, takes up related questions about independent judgement, human relationships and unresolved uncertainty. Its status is important: Microsoft says the draft is not currently being used to train its models. A revised version is intended to guide model development in 2027 and beyond.

For schools, this provides a set of proposed commitments to examine, rather than a guarantee about today’s products. A deployment through Microsoft Teams should not be assumed to inherit safeguards intended for future MAI models.

The useful response is to turn the language into classroom tests. Give a tutor a genuinely ambiguous problem. See whether it acknowledges the ambiguity, helps a learner work through it, or produces a confident answer that closes the discussion prematurely. Then examine whether the learner can explain their reasoning without the tool.

For an AE Engine pilot, a manageable next step would be to bring three things together: a short explanation for teachers and families, a small set of shared test scenarios, and an unaided follow-up activity. Together they could help distinguish an impressive interaction from learning that lasts. If the same team develops the product and evaluates it, that relationship should be explicit.

A promising pilot should leave the school better able to decide what to do next—including changing the approach or stopping. That capacity to make an informed decision is itself worth designing for.