Keep generative AI on the teacher's side of the desk

There's a staff-room question that keeps getting asked as if it has one answer: Is generative AI allowed in our classrooms, or not? Ban it and you ignore what teachers are already doing after hours. Put a chatbot on every student device and you pretend a fluent guessing machine is a tutor. Both moves dodge the only placement question that actually helps: who sits in front of the generative interface — the chat box, the "make me a…" panel, the thing that invents new text on demand?

Our answer is blunt. Keep generative AI on the teacher's side of the desk. Use it to prepare the lesson. Do not use it — and do not accidentally share it — as the thing that answers the kid.

Two different jobs for the same model

The confusion starts because people treat "using AI" as one activity. It isn't. There are at least two jobs, and they have different failure modes.

Job one is production. An adult who already knows the content asks a model to draft structures: a worksheet skeleton, three versions of the same exit ticket, a cleaner layout for a lab sheet, a remixed packet for a sub day, a checklist header that stays consistent across a unit. The output is an artifact. The teacher edits it. Students eventually see a PDF, a printed page, or a slide — not a chat transcript.

Job two is tutoring-by-chat. A student types a question into a live model and gets a finished answer in a confident voice. The model is not handing back a source the kid can argue with. It is producing a paragraph that sounds like knowing. That is a different cognitive contract, and for early learning especially, it is a bad one.

Same underlying technology. Opposite roles in the room. Treating them as the same "AI policy" is how schools end up either terrified of worksheets or casual about chatbots.

Bad default	Better default
Student asks the model for the answer to today's facts	Teacher asks the model for a clearer template to present those facts
LMS task: "If you're stuck, ask the chatbot"	Stuck path: hint from you, a worked example you prepared, or a peer
Share the chat link because export felt slow	Export to PDF / slides / print — then close the chat
Open-ended student chat as "enrichment"	Older students building with AI under supervision — a different activity, designed on purpose (age-appropriate access)

That last row matters for high school: teenagers will encounter models. Supervised building can be the point. Live fact-chat as the default way a nine-year-old — or a junior — finishes a worksheet is still a bad default. Don't use "they'll find it anyway" as cover for putting an oracle on every desk.

Not what this rule is about: assistive tools a student needs under an IEP, 504, or language-access plan (read-aloud, dictation, translation) are not the same as open-ended generative chat answering content for them. This essay is about placement of the invent-on-demand interface — not a ban on accessibility.

Why live chat is a bad default for students

We are not allergic to AI. We run a program about building with it. We are allergic to pretending a chat window is a gentle private tutor for kids who cannot yet reject a fluent lie.

It outsources the thinking that school is supposed to build. Asking a model for the answer to a fact or a procedure you don't know yet skips the struggle that makes the knowledge stick. In early grades that struggle is the curriculum. A tool that collapses it into a polite paragraph is not "personalized learning." It is skipping the rep.

Kids can't catch the model's most dangerous failure mode. Language models are fluent when they are wrong. An adult who knows fractions sees a wrong explanation and flinches. A third-grader sees neat sentences and assumes the neatness is truth. If you wouldn't hand a child an authoritative-looking textbook written by a stranger who sometimes invents citations, you shouldn't hand them an interactive version of the same risk with a smile and a send button.

The interaction is hard to supervise. A packet on a desk has a surface area. A chat thread is private, fast, and easy to hide inside a tab. Our safety posture is built around supervised rooms, not private oracles — sightlines, account gates, adults who can see the work. None of that gets easier when the "activity" is the unsupervised conversation.

It muddies the relationship kids are supposed to have with adults and peers. A model that never gets tired, never says "I don't know," and never sends you to try again with a friend will happily occupy the slot where a teacher, a classmate, or a hard problem should sit. That isn't a neutral efficiency gain. It's a substitution.

Why teacher-side templates are a real hack

None of that means teachers should pretend these systems don't exist. Used upstream of the student, generative AI is good at exactly the boring, high-leverage prep work that eats evenings:

• Turning one solid activity into a stable, easy-to-read template you can reuse.
• Remixing existing content — your old lab sheet, last year's quiz, a district sample — into today's constraints (shorter directions, larger type, a sub-friendly version).
• Producing consistent structures: same checklist pattern, same reflection stem, same "materials / steps / check" layout across a week so kids learn the format once.
• Drafting multiple versions fast — three reading levels of the same prompt, a challenge extension, a scaffolded starter — so differentiation isn't a weekend project.

Picture a Thursday night: you already know tomorrow's science lesson cold. You paste last year's lab sheet into a chat and ask for the same procedure with shorter sentences, a materials checklist at the top, and a third version with sentence starters for students who need them. Twenty minutes later you have three pages to markup with a red pen. The kids never see the chat. They see a clearer lab. That is the hack.

The win is not "the AI planned my unit." The win is rapid, consistent production of materials you still own. The model is a layout and remix engine. You are still the author of record.

Yes, that is still work. The point isn't zero effort — it's moving effort from blank-page formatting to judgment. Teachers are already overloaded; the useful tools are the ones that shrink the mechanical part without stealing the part only a subject-matter adult can do.

That only works if the generative step stays offstage. Students get the worksheet. They do not get the chat that made the worksheet.

Expertise is the quality gate — not a nice-to-have

Here is the part that gets skipped in breathless "10 prompts for teachers" threads, and it is the whole game.

Generative tools work best when the person using them already knows the content. Not "has a standards PDF open." Knows it well enough to spot the wrong answer, the slipped definition, the diagram that doesn't match the procedure, the instruction that is ambiguous on paper even though it sounded fine in chat.

Without that, error patterns cascade. You accept one slightly wrong example. You reuse the prompt. You generate a week's worth of materials from the same broken seed. Suddenly an entire packet is polluted with the same confident mistake, laid out neatly, photocopied for thirty kids. The model did not "save you time." It industrialized a misunderstanding.

The practical loop is narrow and adult:

1. Generate a draft structure.
2. Read it like a hostile editor — facts, sequence, wording, visual consistency.
3. Fix the prompt (or the draft) for the pattern of the error, not just the one typo.
4. Only then freeze the artifact students will see.

If that sounds like work, it is. It is also the same judgment we argue adults need before they supervise kids with these tools at all — you can't teach a tool you've never used. Forty hours of honest personal use teaches you what fluent wrongness feels like in your gut. That gut is what keeps a template helpful instead of contagious.

Asking a model for facts you don't know, then putting those facts in front of early learners, is the failure mode. Asking a model for a clean, stable template to present a lesson you do know — then editing until it's true — is the useful one.

How the chat accidentally reaches kids

Most leaks are not dramatic. They are logistical.

A teacher generates a great packet in a chat window, then shares the chat link because it was easier than exporting a PDF. A well-meaning LMS task says "if you're stuck, ask the chatbot." A classroom account is left signed in on a student laptop. A "teacher tool" with a student-facing chat mode gets turned on for "enrichment." The generative interface arrives on the kid's side of the desk without anyone deciding that was the pedagogy.

The boring safeguard is also the right one: generate → export → teach from the artifact. Print it. PDF it. Paste the final text into the slide deck you already control. Close the chat. Sharing a cleaned Google Doc or PDF you edited is fine — the kids are seeing your materials, not the generative session. If a student needs help, they get a hint from a human, a worked example you prepared, or a peer — not a private oracle.

If your school wants older students building with AI under supervision, that is a different activity with different design constraints. It is not the same as opening live chat as the default way to finish a worksheet. Don't smuggle one inside the other.

This week, on the teacher side of the desk

If you only do one practice loop:

1. Pick a lesson you already know cold.
2. Paste an existing worksheet or lab sheet into your usual model.
3. Ask for structure only — shorter directions, a materials checklist, one scaffolded version. (Starter prompts below.)
4. Read the draft like a hostile editor. Fix the pattern of any error, not just the typo.
5. Export to PDF / slides / print. Close the chat. Teach from the artifact.

Starter prompt A: "Rewrite this worksheet for [grade]. Keep every learning goal the same. Shorten directions to one sentence each. Add a materials checklist at the top. Do not add new facts or answer keys."

Starter prompt B: "Turn this activity into three versions with the same structure: (1) standard, (2) shorter sentences + sentence starters, (3) challenge extension. Preserve my sequence. Flag anything you're unsure about instead of inventing."

District-required tutor pilots still get the procurement test in the next section: if the generative interface sits on the student's desk by default, you know what you're buying — and what you're giving up.

One rule you can defend in a staff meeting

Say it out loud:

Generative AI may help adults prepare materials. It should not be the live tutor answering for students — especially when they are still learning to tell a fluent sentence from a true one.

That rule is pro-teacher and anti-outsourcing at the same time. It lets you use the real hack — templates, remixes, consistent structures — without pretending a chat window is a substitute for teaching. It also gives you a clean test for any new tool demo: Does this put the generative interface on the teacher's desk or the student's? If you can't answer, you are not ready to buy it.

This is a placement rule, not a district policy manual. Buildings will still need their own tool lists and parent communication. Start with the desk sides clear, and the rest gets easier to argue.

If you're a caregiver trying to decode what "AI at school" should look like, start with how we think about safety. If you want a short, grown-up path into using these systems for your own work first — before you decide what belongs near kids — start at for grown-ups. If you want classroom materials that arrive as finished printables rather than as a chatbot, the free printable library is built that way on purpose: the AI may have helped in the kitchen; the kids only get the plated dish.

Keep the chat where the prep happens. Keep the lesson where the learning happens. The desk has two sides for a reason.