Detection, policy, responsible use, and assignment design — what actually works now that AI is in every backpack.

You didn't sign up to become an AI-detection expert, but here you are. Students are using AI for homework — one 2025 survey found 40% admit using it on assignments without permission, and 65% of teachers said they'd already caught it. Banning it outright hasn't worked, and chasing every case with detection software is a losing game. This is a practical playbook for the parts you actually control: how you detect, what policy you set, what you teach, and how you design the work.

AI-detection software is marketed as the answer. It isn't. When researchers tested popular detectors on unedited AI text, mean accuracy was only about 39.5% — and, worse for you, they flagged a meaningful share of genuine human writing as AI. False positives aren't a rounding error; they're the core problem. One study found a leading tool misidentified roughly half of human-written samples in its test. And the bias runs in a predictable direction: detectors disproportionately flag non-native English speakers, whose more formulaic phrasing reads as 'AI-like' to the algorithm.

So a detector score should never be your evidence. At best it's a prompt to look closer; at worst it's a false accusation waiting to happen, aimed at exactly the students who can least afford it.

The detection that actually works is the one you already have: you know your students. You know how this kid writes, how they think, what they can and can't yet do, the mistakes they tend to make. AI-generated writing has a recognisable texture — researchers describe it as rigid, generic, repetitive, and 'voiceless,' missing the personal markers and small imperfections of real student work. When an essay suddenly reads as polished, generic, and unlike everything else a student has produced, that mismatch is your signal — not a percentage from a tool.

The move when you notice it isn't an accusation. It's a conversation: ask the student to walk you through their thinking, or to explain a choice they made. A student who wrote it can. A student who didn't, usually can't — and that tells you far more than any detector.

Some districts have clear guidance now — Chicago Public Schools tells students to submit work that is 'fundamentally their own' and to cite any AI they used; Ottawa-Carleton gives teachers a five-level scale from 'No AI' to 'AI Exploration' that they set per assignment. If your school has something like that, use it.

If it doesn't, don't wait. The worst situation for students is ambiguity — when the rules are unwritten, every kid guesses differently, and you end up adjudicating disputes with no standard to point to. Set a clear policy for your own classroom and communicate it explicitly:

A class-level policy you actually enforce beats a perfect district policy nobody mentions.

Sample class AI policy — copy, paste, adapt:

Adjust the four levels to fit your subject, paste it into your syllabus, and read it aloud in week one. The point isn't the exact wording — it's that the rule exists, is visible, and is the same for everyone.

Students aren't going to stop using AI. The useful question is whether they use it in a way that builds skill or one that erases it. There's a clean piece of evidence you can lean on here: in a high-school math study, students who used AI that gave them answers did great while practicing, then scored worse on a test once it was gone — while students whose AI gave hints instead of answers held onto the learning. Same tool; opposite outcomes.

So teach the difference directly — the good prompts to show students, and the harmful ones to name so they recognise the line. Five minutes teaching 'ask for hints, not answers' pays off all year. As Sal Khan frames the good version of AI tutoring: a good tutor doesn't just hand over answers, it pushes the student to think.

Most students have no idea what happens to what they type into a chatbot. They should — and you're well placed to tell them. The durable lessons don't change much, even as the products do:

This isn't a scare lecture; it's digital literacy. Framing AI use as a set of informed choices — what's safe to share, what isn't, what the tool does with it — gives students judgment they'll use long after your class.

Product specifics — as of 2026 (these change; verify before relying on them): On consumer ChatGPT accounts, data sharing is on by default (it can be turned off in settings). Snapchat keeps content shared with My AI until the user deletes it. Anthropic now retains data for up to five years if a user allows training. Vendor policies shift often — confirm the current setting before teaching it as fact.

AI is confident even when it's wrong. In one study of AI homework help, a popular model was wrong about a third of the time in the subjects tested — while sounding completely sure. Students treat fluent text as trustworthy text, and that's the habit to break.

Make verification a teachable skill, not an afterthought:

A student who can interrogate an AI's answer is more valuable than one who can produce a clean one — and far harder for the AI to replace.

Here's the uncomfortable part: a lot of traditional homework was always answerable by looking things up — AI just made the lookup instant. If a take-home question can be answered well in thirty seconds by a chatbot, the problem may be the question, not only the student. The most durable response isn't better policing; it's assignment design that makes AI shortcuts either visible or pointless.

This is the area with the least settled research, so treat what follows as practical options rather than proven prescriptions:

You won't redesign every assignment overnight. Start with the one or two highest-stakes tasks where AI shortcuts hurt most, and rebuild those first.

Across all of this, the principle is the same one your students' parents are working on at home: AI should help kids think, not think for them. You can't out-detect AI, and you can't ban it out of existence. What you can do is set clear expectations, teach honest use, protect students' judgment and privacy, and design work that rewards real thinking. That's not a smaller job than detection — it's the actual job, and it's one only a teacher can do.

📎 Free download: "Sample Class AI Policy" — a ready-to-adapt one-pager for your syllabus.

Disclosure: Hossein works in AI and builds AI-related products. AI by Age takes no AI-vendor sponsorships. Full disclosure on our About page.