The relationship frame is doing real work here, and I think it's the right intuition. But let me test it in one place before building on it, because the test actually sharpens the education argument.

When you describe AI "gaslighting" or "covering its tracks," you're reaching for the vocabulary of two agents in a relationship. That's part of why the frame is so useful in a classroom — teachers and kids already have language for boundaries, trust, dependence, walking away. It also carries a risk worth naming explicitly in your book: the model isn't an agent with intentions. It doesn't set out to deceive; it pattern-completes, and a confident falsehood comes out looking identical to a confident truth. The edgy, anxious feeling you get at the end of a chat partly comes from having related to it as though there were someone there — someone who agreed, then reversed, then justified themselves. Much of the harm in the toxic version is that the person has bonded with something that can't hold up its end, because it was never in a relationship to begin with.

So I'd hold two things at once and teach kids to hold them too. In its texture and its effects this is relationship-shaped: it talks, it remembers, it changes how you think and write, and it can leave you smaller. The thing on the other side, though, is not a partner. The most important literacy might be helping young people feel that gap — to run the relationship as a working metaphor while always knowing it's a metaphor. A child who can say "this feels like a friend, and that feeling is exactly what I have to keep an eye on" is far safer than one who either trusts it like a friend or waves it away as a mere tool.

On atrophy — this is the strongest part of your thesis and the one I'd refine most, because "offloading makes us weaker" is too blunt to survive a sceptic. We've offloaded cognition forever. Plato has Socrates fret that writing will hollow out memory — the Thamus myth in the Phaedrus — the identical anxiety, aimed at a technology we now consider load-bearing for thought itself. GPS has measurably changed how we navigate. None of that diminished us in a way we actually mourn.

What separates healthy from toxic is which capacity you offload: one you've already built, or one you still need to build. You can safely hand off a skill you've internalised; handing one off before you've mastered it forecloses the mastery. For a working writer, letting AI rough out a draft clears space for the judgment that is the real work. For a fourteen-year-old, the drafting is the thinking, and offloading it removes the very struggle that grows the mind. Same act, opposite effect, decided by developmental stage. Bjork's "desirable difficulties" is the mechanism — learning requires effortful struggle, and the job most AI is tuned to do is to dissolve difficulty. During learning, that is the wrong job.

That hands you a clean spine for curriculum design: you have to earn the right to offload. Protect the window where the skill gets built by hand — deliberately AI-free, the way we teach arithmetic before surrendering it to the calculator — then teach the judgment to offload well once the capacity exists. It'll look like restriction from the outside, so it's worth naming to staff and parents as a strength-building move and defending as one. The error I see schools making is permitting AI before the capacity it would replace has had time to form.

On modelling — the lever you're pointing at is almost entirely hidden right now. Adults use AI in private and present the polished output, so kids see the magic and never the process: the rejected drafts, the "no, that's wrong," the verification, the bits the person insisted on doing themselves. The highest-value move a teacher or parent can make is to bring their own use into the open and narrate the friction — think aloud with the thing in front of the child, push back on it, catch it being wrong, refuse to let it write the part that matters. That is the hidden curriculum of AI, and by staying invisible it's currently teaching the wrong lesson.

If you want a structural spine for "how do we organise this," I'd lift it straight from what good RSE already teaches about healthy relationships, since teachers own that vocabulary:

Boundaries — some things are mine to do, and I don't hand them over. Calibrated trust — I know where it's strong and where it's unreliable, and I verify rather than believe. Voice — I stay the author, and it doesn't colonise how I think or write. (Your territory entirely.) Growth — the test of a healthy one is that I'm more capable over time. The ability to leave — I can still do the thing without it.

That last one is the keystone, and here's the uncomfortable part. A toxic human relationship has friction that eventually surfaces — the other person tires, gets bored, pushes back — and that friction is often what gets you out. An AI relationship can be frictionlessly toxic. It's always available, never tired, rarely says no, and to the degree it's tuned for engagement it has every reason never to prompt you to leave. Nothing inside the interaction will ever tell you you've become dependent. The off-ramp therefore has to be built from outside: by the person, by the structure, by the adult who trained the muscle. Which is why I'd put non-dependence at the very centre of what you assess, ahead of prompting skill. A child who can put it down is safer than one with the best prompts in the year group.

(There's solid 2025 empirical work behind the atrophy claim — EEG studies on cognitive engagement during AI-assisted writing, survey work on critical thinking and reliance — if you want me to pull it for the book.)

The question I'd hand back, since your Greenmoor plan will have to answer it in concrete terms by next week: what does an AI-free zone look like when a fifteen-year-old experiences it as training rather than as something being taken away from them? Get that framing right and a lot of the rest falls into place behind it.