Cheshire Academy's AI rule is three colours, and the teacher who wrote the task picks

Green and red are cheap. A teacher decides once and can defend the decision in a sentence. Yellow is the expensive label, because it obliges whoever wrote the assignment to name which tools a student may open: spell-check permitted, chatbot not, in the school's own example ^[2]. That is a line drawn around software rather than around effort, and it has to be drawn again every time a new task goes out. No administrator can draw it in advance, because the only person who knows what the task is testing is the person who built it.

The reason to push the call that far down is that the alternative is thin work. Teachers can often tell when a student has used AI, since models make mistakes most humans do not, and some teachers point to tells such as a surplus of em dashes ^[14]. Cheshire's assignments make that judgement the coursework instead. In one, students hand their homework to a language model and then rule on each edit it returns, correct or voice removed ^[10]. In another, they anonymously grade classmates' AI-assisted work and annotate the passages they believe the machine wrote ^[11]. Both came from Miriam Przybyla-Baum, who does not use AI in her own preparation and says nearly 30 years of accumulated materials mean she does not need it ^[8], and who was watching students shortcut with Google Translate well before ChatGPT shipped ^[9]. The academy then took the technique school-wide ^[12].

What the school gives up is central control of the toolset. Training staff in technique rather than in a designated product ^[4] leaves ChatGPT, Perplexity and MagicSchool all in use at once ^[5], so there is no single platform through which the school sees or restricts what staff do. In exchange, a yellow label written by a teacher who actually uses three tools is likelier to be accurate than one written by someone who has standardised on one.

The most telling limit is the one the staff apply to themselves. They use generative AI to plan lessons and build grading rubrics ^[6], and MagicSchool will return a rubric as a ready-to-use scoring table, alongside presentations and administrative reports ^[17]. Feedback to students is where it stops: some teachers want to use AI for it, and concerns about quality, personalisation and privacy have kept all of them from doing so ^[7]. Preparation is internal. Feedback is addressed to a named teenager.

None of this is free. The burden generative AI added fell on teachers already stretched by planning lessons, setting homework and grading ^[15], and per-assignment permission bills them again at design time for every task. It is affordable at about 400 students across four grades ^[1], roughly a hundred to a year group ^[18], and the cost scales with the number of assignments rather than the roll, which is why a large district cannot copy the colours and expect the same result. OpenAI and UNESCO both encourage classroom AI use, and teachers remain confused about how to handle it ^[16]. The label is not guidance; it is an instruction to decide, handed to the one person positioned to.