At Cheshire Academy, a private school in Connecticut with about 400 students, there is no single artificial intelligence policy. Teachers use different tools — ChatGPT, Perplexity, MagicSchool — and integrate them into lessons as they see fit. It is not chaos, although at first glance it may look that way: it is a diffuse laboratory in which the school chose to train staff on general prompting techniques and on the limits of LLMs, rather than imposing a single platform.

The choice has a precise logic. Generative chatbots arrived in classrooms before institutions understood how to govern them. Responses can be wrong or biased, and even the generated text has telltale signs. Forcing everyone to use the same tool would have created vendor dependence and ignored the fact that every subject has different needs. Better, in the school's view, to prepare teachers to choose.

The result is a range of practices. Some use AI to plan lessons and create grading rubrics. Others, like French teacher Miriam Przybyla-Baum, do not need it for preparing materials — after nearly thirty years of teaching she already has a consolidated archive — but use it as a critical mirror for students. In one exercise, students have a large language model edit their homework, then distinguish correct edits from those that erase their voice. In another, they anonymously grade each other's AI-assisted assignments, annotating which parts they consider artificially generated.

This strategy has produced a simple but effective mechanism: the traffic light. Green means AI fully allowed; red an absolute ban; yellow authorizes only certain tools, such as a spell-checker but not a chatbot. It is a pragmatic response to the control problem: you cannot police every take-home assignment, but you can make the boundary explicit.

The hidden cost of freedom

From the perspective of someone observing LLM adoption in organizations, the Cheshire Academy case is instructive. The freedom granted to teachers places an additional burden on them: figuring out which tool to use, checking its outputs, adapting assignments. Not everyone feels comfortable using an LLM to generate student-facing text, whether because of doubts about pedagogical effectiveness or concerns about accuracy. The school has so far blocked the use of AI for personalized feedback for exactly these reasons: quality, personalization and privacy remain unresolved issues.

This tension is structural, not educational. Specialized platforms like MagicSchool concentrate quizzes, worksheets, rubrics and lesson plans in a single package. But the free version has limits and the paid individual plan costs just under 100 dollars per year: a cost that, multiplied across an institution, introduces budget considerations which often translate into the use of general-purpose chatbots. These work well for administrative tasks but are not designed for teaching.

The on-premise deployment parallel

For those evaluating on-premise deployment, the school case offers a parallel. The choice between a vertical platform and general-purpose tools reproduces the trade-off between control and speed of adoption. A specialized platform promises coherence and governance but binds to an ecosystem and a licensing model. General-purpose chatbots offer immediate flexibility but shift verification and security responsibility onto users. In both cases, the critical point is not which model is adopted but who guards the boundary between productive and lazy use.

Cheshire Academy is trying to address this point with a pilot program called Student AI Council, in which students create content and lead discussions on healthy AI use. It is a signal: while vendors add school features with mixed results, the institution shifts attention from technology to culture of use. It is not a final solution. But it recognises that the real bottleneck, in the classroom as in an enterprise, is not model power: it is the ability to establish shared rules before habit becomes dependence.