AI Aced My Homework. Here's How I Rebuilt My Classroom.
In this article
I'll write this article directly as prose — no code or file operations needed for this task.
Bottom line: When ChatGPT 5 started completing my year-10 history assignments at a B+ level in under 90 seconds, I stopped assigning take-home essays entirely and moved 80% of writing into class time under direct supervision.
The shift — inspired by a similar move documented by teachers on Hacker News in mid-2026 — trades some coverage of content for verified, in-person thinking.
Grades on in-class work dropped initially, then recovered as students adjusted to writing without AI as a crutch.
If your students have laptops and an AI account, take-home essays are no longer a reliable measure of what they know.
My year-10 student Marcus turned in an essay on the causes of World War I that was better than anything he'd written all year.
Clean thesis, tidy paragraph transitions, a quote from a primary source I didn't recognize. I gave it a B+ and moved on to the next stack.
Three weeks later I ran the same prompt through ChatGPT 5 myself, out of idle curiosity, and it spat out something almost identical in under two minutes.
I sat there for a while just staring at the screen. Not angry, exactly. More like the floor had quietly moved six inches to the left and nobody had told me.
The Real Tension: I Couldn't Tell Anymore
Here's what's actually broken, and it's not "kids are cheating." Kids have always found shortcuts — CliffsNotes, older siblings' essays, Wikipedia summaries dressed up in different words.
What's broken is that the shortcut is now indistinguishable from the real thing, and it's free, instant, and available on every phone in my classroom.
I tried the AI-detection tools first, like most teachers did in early 2026.
Turnitin's detector flagged one of my strongest writers as 94% AI-generated — she'd written every word herself, she just writes in clean, structured sentences because I taught her to.
It flagged a genuinely AI-assisted essay from another student as 12% likely. The tools weren't just imperfect, they were confidently wrong, which is worse than not having them at all.
So I stopped trying to catch anything. That's the part that surprised me most — the fix wasn't better detection, it was giving up on detection altogether.
Once take-home writing stopped being verifiable, I had to stop grading it as if it were.
The tension underneath all this isn't really about AI.
It's about the fact that most of what I assessed for a decade — a finished five-paragraph essay, completed alone, at home — was always a proxy for thinking, not thinking itself.
AI just made the proxy collapse overnight. A similar reckoning is playing out in workplaces where AI-assisted output looks identical to the real skill until someone asks a follow-up question.
What's Actually Working: Moving the Thinking Back Into the Room
The single biggest change I made was converting roughly 80% of my writing assignments from take-home to in-class, done on paper or on locked-down Chromebooks with internet disabled.
It sounds almost embarrassingly low-tech. It is. It also works, because you cannot outsource thinking you have to do while I'm standing four feet away.
A few specific things that have held up over a full term:
Timed, scaffolded drafting in class. Instead of "write an essay on the Treaty of Versailles, due Friday," I now give 35 minutes of in-class time with an outline template, a source packet, and zero devices.
Students draft the body paragraphs live. I collect it before they leave.
It's not glamorous, but I can see the handwriting shake when someone's stuck, which tells me more than any polished paragraph ever did.
Oral defense of written work. For anything longer that I do let students take home — a research project, say — they now have to sit with me for five minutes and explain a choice they made in their own paper: why this source, why this structure, why this conclusion.
A kid who used AI to generate the whole thing cannot answer "why did you put this paragraph second" convincingly. A kid who wrote it, even badly, always can.
Process documents over polished final products. I now grade the messy stuff — annotated sources, a scratch outline, a paragraph that got crossed out and rewritten — as heavily as the final draft, sometimes more heavily.
Version history in Google Docs has become genuinely useful here; a document that appears fully formed in one paste is a flag worth a conversation, not an accusation.
Teaching AI as a tool, openly, in select assignments. For one unit a term, I now explicitly allow ChatGPT 5 or Claude for a specific, narrow task — generating counterarguments to test against, say — and grade the student's critique of what the AI produced.
This isn't surrender. It's closer to teaching kids to use a calculator well instead of pretending long division by hand is still the point.
None of this eliminated AI use. Kids still use it at home for other subjects, and I have no illusions they've stopped.
What changed is that I stopped grading things I couldn't verify, and started grading things I could watch happen.
The Mental Model: The Three-Question Assessment Filter
I built a simple filter I now run every assignment through before I hand it out, and it's made planning genuinely faster, not slower.
1. Could a current AI model do this entire task from the prompt alone, with no access to my class discussions? If yes, either redesign it or move it into a supervised setting.
A generic "analyze the themes of Macbeth" prompt fails this test instantly.
"Analyze the themes of Macbeth using three specific moments we discussed in Tuesday's class" mostly survives it, because the AI wasn't in the room.
2. Can I see the thinking, not just the output? Drafts, annotations, oral explanation, in-class time — any of these count. A final PDF alone does not.
3.
Does this assessment reward a process I actually want to reinforce? If the honest answer is "no, it just rewards producing 800 competent-sounding words," it's not worth keeping even if AI weren't in the picture.
I call it the Visible-Thinking Filter, and it's stuck with a few other teachers at my school precisely because it's not about AI policy at all — it would have improved my assessments even in a world where ChatGPT didn't exist.
It just took AI breaking the old system for me to notice how much of it was theater.
Honest Failure Modes
I got some of this wrong at first, and it's worth saying so plainly.
My initial instinct was to swing hard toward surveillance — locked browsers, oral quizzes on every single assignment, a suspicious eye on any essay that read too smoothly.
Kids noticed, and it curdled the classroom mood fast. A parent emailed me, reasonably annoyed, because I'd questioned her daughter's authentic, well-written essay simply for being well-written.
Treating every strong writer like a suspect is its own kind of failure, and it punishes the kids who were never the problem.
The oral-defense requirement also backfired for my more anxious students, who know the material cold but freeze up explaining it verbally on the spot.
I've since made it optional to submit a short recorded voice note instead of a live conversation, which solved most of it.
And the in-class-only writing came with a real cost: I cover less content than I used to, because 35 minutes of supervised drafting eats time I'd previously spent lecturing.
I'm not pretending that trade-off is free. Some units are thinner than they were two years ago, and I don't yet know if that catches up with these kids by the time they hit their exams.
The Quiet Win
The quiet win isn't that I "beat" AI or found a clever workaround.
It's that Marcus — the kid whose suspiciously good essay started all this — actually wrote a genuinely strong paragraph in class three weeks later, out loud, under pressure, with nothing to lean on but what he'd actually understood.
It wasn't as polished as his AI-assisted one. It was his.
One thing to try this week: take your next writing assignment and convert just the introduction and one body paragraph into a 20-minute in-class task, done by hand, no devices, collected before the bell.
Don't redesign your whole course. Just see what you can actually see.
Has your classroom — or your kid's — already crossed this line, or are you still grading essays you can't quite trust?
