Ask a staff room how students use ChatGPT and you'll hear one word: cheating. Ask the students and you get a much stranger, much more interesting answer.
The UK's Higher Education Policy Institute has been tracking this, and their numbers move fast: 53% of undergraduates used generative AI for assessments in 2024, then 88% in 2025. Whatever your policy says, that's the installed base you're governing. But the headline number hides the part that actually matters, which is what they're doing in there.
The same tool, four different students
Watch real usage (we do, on school devices, it's the product) and the sessions sort into recognizable shapes:
The first student pastes a photosynthesis diagram and types "explain this like I'm dumb, my teacher went too fast." That's a tutoring session. It's the kind of help wealthy families have always paid for by the hour.
The second asks for five practice problems on integration by parts, harder each time, with worked answers after each attempt. That's studying. Arguably better studying than the textbook offers.
The third writes two paragraphs, pastes them in, and asks "does my argument make sense here?" That's feedback. A writing center visit at midnight.
The fourth pastes the assignment prompt and types "write 800 words on this, make it sound like a high schooler." That's the one your policy was written about.
In the survey data, students confirm this spread: the most common uses they report are explaining concepts and summarizing material, with a smaller share admitting to submitting generated text. Four behaviors, one tool, one domain name in your network logs.
Every blunt policy punishes someone innocent
Now run those four students through the policies schools actually deploy.
Ban the domain, and you took the free tutor away from the first kid, who by the way just moves to a phone. HEPI's own finding is that students without other support lean on AI hardest, so the ban lands heaviest on the kids with the least backup. Allow everything, and the fourth kid graduates on outsourced essays. Run detectors on the output, and you're guessing about finished text with a tool famous for flagging honest writers.
Notice what all three approaches have in common: they operate on the tool, because the tool is all they can see. The behavior is invisible to them. And the behavior was the entire point.
Govern the sentence, not the site
The uncomfortable truth for policy writers is that "is ChatGPT allowed" was never a real question. "Explain this to me" and "do this for me" are different acts that happen to share a URL. Any rule that can't tell them apart is guaranteed to be wrong in one direction or the other, every single day.
This is exactly the distinction Themisto EDU is built to make. On school devices, it reads the intent of the AI interaction as it happens: tutoring, practice, and feedback go through; produce-my-assignment gets blocked before it's sent, and the event lands in a dashboard a human can review. The first three students keep their midnight tutor. The fourth one has a different night than planned. Teachers stop guessing from vibes and detector scores, because they have the actual record.
Nine in ten students are already in the room with this thing. The question left is whether the adults can see what's happening in there. Our free classroom snapshot takes about a minute and gives you the starting numbers for your school. It's a better opening move than another staff meeting about the word "cheating."