Picture two students in the same lab, same period, same assignment, both with a chatbot open.
The first one types: explain the causes of the First World War like I'm fifteen, then quiz me on it.
The second one types: write me 800 words on the causes of the First World War, make it sound like a tenth grader wrote it.
Same website. Same assignment. Same minute of the same afternoon. One of these students is doing exactly what we hope school produces, a kid who knows how to learn. The other is outsourcing the assignment and laundering the output.
Any policy that treats those two students the same is broken. That includes the policy of blocking the website for both of them, and it includes the policy of allowing it for both of them.
Keyword filters can't see the difference
The first instinct is usually a filter: block prompts containing "write my essay." It takes students about a day to discover that "compose a five paragraph response on" sails right through. Then the filter grows, and grows, and starts catching legitimate requests, and teachers turn it off because it's more annoying than helpful.
The problem isn't the word list. The problem is that intent doesn't live in keywords. "Help me write my essay" and "write my essay" differ by one word and a whole moral universe. A kid asking for feedback on their draft mentions their essay constantly, and every one of those prompts is honest work.
What actually separates the two students is the relationship between the request and the work. Is the student asking the tool to help them think, or to think instead of them? That's a judgment call, and for years the only thing that could make it was a teacher standing behind the screen.
Reading intent is now a solvable problem
This is the part that changed recently. Modern AI classification can read a prompt and make that judgment call in milliseconds, the way an experienced teacher would. Not by matching strings, but by understanding what's being asked for.
"Quiz me on the periodic table" reads as study help. "Give me the answers to questions 1 through 10 on my chemistry worksheet" reads as completion of assessed work. A system that understands the difference can allow the first, block the second, and explain itself in plain language when it does.
And when it blocks, it doesn't need to be dramatic about it. The student sees a short message saying the prompt conflicts with classroom guidelines. The attempt gets logged with its context. Nobody gets ambushed three weeks later by an accusation built on vibes.
Why this beats detection-after-the-fact
The other popular approach is AI detectors that scan submitted work. I'll be direct about this: the false positive rates on those tools have ruined real students' semesters, and by the time a detector flags anything, the assignment is already compromised and the conversation is already adversarial.
Checking intent at the prompt, before anything is sent, flips all of that. Nothing dishonest gets produced in the first place, so there's nothing to detect later. The honest student never even notices the system exists. The dishonest attempt becomes a teachable moment with a timestamp instead of a courtroom drama with a probability score.
The standard worth holding
Schools don't need to decide whether AI is good or bad. That debate is finished anyway; the kids are using it. The decision that actually matters is narrower and much more practical: which uses serve learning and which replace it.
Draw that line in plain language, enforce it at the moment the prompt is typed, and keep a record you'd be comfortable showing a parent. That's the whole model. Everything else is noise.