I want you to imagine a school where the district doesn't adopt a curriculum. Instead, every teacher just... picks their own. Ms. Alvarez orders a textbook she found online. Mr. Chen builds his own packet from things he liked in grad school. The teacher next door uses whatever's left in the supply closet. Nobody's textbooks match. Nobody's pacing lines up. A kid who moves from third period to fourth period is, functionally, changing schools.
You'd never allow that. No principal would. No superintendent would sign off on it. And yet — walk down the hallway right now, today, and ask five teachers how they're using AI with students. You'll get five different tools, five different setups, and five wildly different qualities of output. Some teachers have quietly built something good. Most haven't. And nobody at the district level is treating this the way they'd treat a textbook adoption.
That's the problem. Here's why it's harder than it looks — and here's what the research says we should actually do about it.
Why the same AI isn't the same AI
An AI model doesn't know your students. It doesn't know your rubric, your unit plan, or the three misconceptions your kids keep tripping over. Unless you give it that information, it's guessing — fluently, confidently, and often wrong. Feed it that information, and something different happens.
Researchers tested this directly. In a study generating curriculum-aligned math questions for secondary schools in Malaysia, the same underlying model was run two ways: once with a plain prompt, and once fed the actual curriculum documents, teacher notes, and yearly teaching plan — a technique called retrieval-augmented generation, or RAG. The grounded version won, and it wasn't close. The documents-in-hand version produced questions with meaningfully better curriculum alignment and factual accuracy than the version working from a prompt alone (arXiv, 2508.04442).
Translation: the model wasn't smarter. It was better informed. That's the whole ballgame.
Two more recent, real-classroom examples make the same point. CyberScholar, tested with 143 students and five teachers across grades 7–11, feeds a writing-feedback model the teachers' own rubrics, materials, and exemplar essays — not a generic prompt. Students said the feedback felt immediate and actually matched what their teacher had been asking for, and they used it to sharpen organization and elaboration as they revised (arXiv, 2605.17055). And a team building an AI tutor plugin for Moodle grounded it in a teacher's own course materials instead of the model's general knowledge, and measured "faithfulness" scores up to 0.97 — meaning the tutor stayed remarkably close to what was actually taught, instead of drifting into plausible-sounding nonsense (arXiv, 2605.06963).
One rigorous, controlled study. Two live tools built this spring. Same finding, three times over: context isn't a nice-to-have. It's most of the outcome.
A necessary pump of the brakes
Here's where I need to steelman the skeptics, because they're right about something important: none of this makes AI a replacement for a teacher. Even the best-grounded systems in these studies still had a human reviewing the output. The Moodle team built in a human-in-the-loop workspace on purpose — because a well-fed model can still get things wrong, and because tutoring is a relationship, not a transaction.
So let's be honest about where this technology actually belongs right now. It's a genuinely strong tool for the adults in a building — teachers building materials, coaches reviewing instruction, administrators managing curriculum. It is not, yet, a good stand-in for a teacher sitting across from a student. Kids deserve a human in that seat. That's not a hedge. That's the point.
But — and this is the part worth sitting with — the lesson about access to context doesn't disappear just because the use case shifts from students to adults. It applies just as much. Maybe more.
So what do we actually do on a Monday morning?
Go back to the textbook. A district doesn't let every teacher order their own because consistency matters, quality control matters, and a kid shouldn't get a worse education because of which room they walked into. We already know this. We built entire adoption committees around it.
We're about to do the opposite with AI — quietly, by default, one classroom at a time — unless somebody decides not to.
So decide not to. Concretely:
- Build shared, secure access to the same underlying files and architecture across a department or district. Every section of Algebra I should be drawing on the same grounded system, not whatever one teacher managed to cobble together over a weekend.
- Ask who vetted what's feeding the AI. You wouldn't adopt a textbook nobody reviewed. Don't adopt a grounding source nobody reviewed either.
- Treat a teacher's homemade AI tool as a pilot, not a policy. Learn from it. Don't scale it by accident.
The research already answered the question of whether grounding matters. It does. Every study points the same direction.
What it can't answer is whether your district builds that grounding once, well, and centrally — or leaves it to chance, teacher by teacher, hallway by hallway.
You wouldn't let every teacher pick their own textbook. Don't let every teacher pick their own AI.
This is part of Teaching in the Age of AI, a weekly digest of research and ideas for educators navigating AI in the classroom. Subscribe to get each week's post.
.png)