This is where we are in September 2026. Eighty-eight percent of students are using AI for schoolwork. Seventy-seven percent of them have received zero formal training on how to do it well.
The numbers come from Microsoft’s 2026 AI in Education Report. The University of Leicester just finished rolling Copilot out to 21,000 students and 4,000 staff in the first week of September. Google is running an all-day AI educator training on September 19. Every institution I read about is adopting AI infrastructure faster than it can train students to use it.
The real problem is the gap between deployment and literacy.
When a tool reaches this level of adoption (when it’s in students’ pockets, in their practice, in their expectations of what school should enable), the old model of “we’ll teach you this later” stops working. Your students already have access. What they need is guidance on what it can do well, where it fails, how to verify its work, when it’s appropriate, and when it’s a substitute for thinking instead of a tool for extending thinking.
Right now, most students are learning this from TikTok, from their friends, and from trial and error on assignments that count. Not from adults who understand both the tool and the learning.
We’ve seen this before, with Wikipedia. When Wikipedia first became ubiquitous, schools did what institutions do: they banned it. When that failed, they pivoted to “use it but cite it properly.” Twenty years later, we landed on something more sensible: Wikipedia is a starting point for research, useful for context and citations, flawed in ways you should understand and verify. The lesson was eventually that we teach students how to read it, not how to avoid it.
But we took two decades and tens of thousands of badly-evaluated essays to get there. With AI, the timeline is compressed. Students don’t have that time to figure this out through institutional bumbling. And neither do teachers.
The design question isn’t whether to add AI to everything. It’s what this tool is useful for in this context, and what conversations students need with it and with you to use it well.
Teach them by demonstration: show what the tool does well and where it fails. Have them use it badly on low-stakes work so they understand its limitations through experience, not warning. Make those mistakes visible and discussable in the classroom, not buried in a grade.
For a DBQ, that might mean: “Use Claude to brainstorm your argument. Show me the brainstorm. Tell me what you changed and why.” For a math problem set, it might mean: “Solve this without AI. Then use Claude to check your work. If Claude’s answer differs from yours, figure out why before you submit.” For a research paper, it might mean: “Use AI to generate potential sources, then verify every citation against the original and explain why each one matters to your thesis.”
In each case, the tool isn’t doing the thinking. The student is. The tool is doing what it’s supposed to do: extending capacity, prompting reflection, catching mistakes that would otherwise go invisible.
The training gap creates a specific kind of harm, though. When students use AI without understanding its limits, they trust it in ways that break down. The risk: a student uses Claude to structure an argument, accepts it without question, and submits it. That’s outsourcing judgment. The right move: a student uses Claude to brainstorm, then actively decides what works and what doesn’t. Same with code — the risk is using it without understanding; the right move is generating and then examining it critically. These aren’t neutral choices. They carry real costs.
And right now, seventy-seven percent of students are making those moves unsupervised.
The school can’t prevent students from using AI. That ship has sailed, and blocking it would be both futile and harmful. Students need to know how to work alongside these systems; that’s a basic literacy now. The school’s job is to collapse the gap between adoption and understanding. To make sure that when a student uses Claude on an assignment, he’s doing it thoughtfully and knows why, not just because the tool was available and it worked.
That takes deliberate curriculum work. It takes teachers who understand both AI and their subject. It takes conversations at the course level, the department level, and the school level about what responsible use looks like in different contexts. It takes willingness to say, “We’re going to teach this out loud instead of pretending you don’t know it exists.”
The alternative is doing what we did with Wikipedia: spending years pretending the problem isn’t there while students figure it out poorly. We don’t have that time anymore.
If you’re thinking through how to build AI literacy into your curriculum — how to teach students to use these tools well without letting the tools do their thinking — my book goes deeper. The AI Doesn’t Know Your Students is available on Amazon and at shouldiuse.ai/book.
Takeaway for Teachers: This week, ask a student how they’re using AI in their schoolwork. Don’t accuse; just listen. You’ll learn what the gap looks like in your own classroom. Then build one assignment where they use the tool in front of you, and you talk through what it did well and where it failed. That conversation is where real literacy begins.
David Jacobson is a high school history teacher. He writes about AI, education, and the messy intersection of the two at shouldiuse.ai.
