Learning how to detect AI-written homework starts with an uncomfortable truth: no tool can prove a student used AI. Detectors guess, and they guess wrong — in both directions. A student who wrote the essay themselves can be flagged. A student who copied an AI answer can pass clean. If you accuse a student on the strength of a detector score alone, you will eventually accuse the wrong student.
This guide shows what actually works: the signs to look for, why detectors are only one weak signal, and the conversation that tells you far more than any score — followed by what to do next, and how to set homework AI cannot do in ten seconds.
What you need before you start (2 minutes)
Have these three things ready before you look at the work:
- The work itself — the essay, the answers, the file with its history if you have it.
- What you know about the student — their usual level, their usual voice, and how they write in class. You are the baseline no detector has.
- Your school’s AI policy — what counts as allowed use, what counts as misconduct, and what the process is. If you do not know the policy, find out before you accuse anyone of anything.
If your school has no AI policy yet, that is the first problem to fix — before any detector.
Step 1: Look for the signs (5 minutes)
Read the work once, looking for these patterns. One sign means nothing. Three or four together mean it is worth a conversation.
- Too perfect, too generic. Flawless grammar, perfect structure, and not one specific detail from your lessons, your class, or the student’s own life. AI writes clean, empty prose.
- A sudden jump in level. The student writes simple sentences in class and submits an essay with advanced vocabulary and complex sentences at home. A jump is possible — but it is a reason to ask, not a verdict.
- No personal detail. The essay never mentions anything that happened in class, in the textbook you use, or in the student’s experience. AI cannot know those things, so it writes around them.
- Inconsistent voice. The introduction sounds like a university essay, the middle sounds like the student, and the conclusion sounds like a textbook. Mixed voices usually mean mixed authorship.
- Wrong details, stated confidently. A quote that does not exist, a date that is wrong, a fact from a different topic. AI invents facts and states them as if they were certain.
- Perfect spelling, shallow ideas. The grammar is flawless and the argument goes nowhere. Human students usually have it the other way round: real ideas, messy grammar.
None of these proves AI use. A strong student can write a perfect essay. A weak student can have a good day. The signs tell you where to look — they are not the verdict.
Step 2: Do not trust detectors (5 minutes)
AI detectors look at a text and guess whether a human or an AI wrote it. Here is what they do not tell you:
- False positives are real. Detectors flag human writing regularly — and they flag non-native English writers far more often than native speakers. A student writing in their second language is the most likely student in your class to be falsely accused.
- False negatives are real too. A student who edits an AI draft, or asks AI to “write it like a 14-year-old”, will often pass clean.
- A score is not proof. “87% likely AI” is a guess about a text, not evidence about a student. No detector company will stand behind its score as proof of misconduct — and neither should you.
Use a detector, if your school provides one, as one weak signal among many. Never as the only evidence, and never as the reason for an accusation on its own. If your school’s policy says a detector score is enough, that policy needs updating — and this guide is a good reason to raise it.
Step 3: Talk to the student (10 minutes)
A five-minute conversation tells you more than any detector. Ask the student to explain their work — in person, calmly, and without accusing them first:
“Can you walk me through how you wrote this? Show me your notes or your draft. Explain this paragraph in your own words — what did you mean here?”
Then listen. A student who wrote the essay can usually explain it, show you a draft, and talk about the hard part. A student who copied it usually cannot explain the argument, cannot show any working, and gets vague exactly where the essay is most polished.
Three rules for the conversation:
- Ask, do not accuse. Start with curiosity, not with a verdict. “Help me understand how you did this” opens a conversation. “I know you used AI” closes it.
- Ask about the process, not the product. Notes, drafts, and false starts are hard to fake. A finished essay is easy to copy.
- Give them a way back. If they admit it, the conversation becomes about learning, not punishment — and that is the outcome that actually helps them.
Step 4: Decide what to do — and redesign the homework (10 minutes)
What happens next depends on your school’s policy. In general, the response should fit the situation:
- First time, low stakes: ask the student to redo the work, in class, with their notes. The redo is the consequence — and the learning.
- Repeated, or high stakes: follow your school’s process — involve the head of department, talk to the parents, and document what happened. Do not improvise a punishment on your own.
- If you are not sure: say so. “I can’t tell from this work alone” is an honest answer, and it is better than a wrong accusation.
Then fix the homework itself. The best defence against AI-written homework is homework AI cannot do:
- Ask for personal experience, class discussion, or something that happened in your classroom.
- Set work to be done in class, on paper, in a fixed time.
- Grade the process, not just the product — notes, drafts, and a short explanation of the argument.
- Ask students to explain their work out loud, even for one minute.
If a homework task can be completed by pasting it into an AI tool, students will paste it into an AI tool. That is not a student problem. That is a task problem.
A complete example
A Grade 9 student submits a history essay that is far better than anything they have written in class. It is flawless, generic, and it quotes a source that does not exist in the textbook.
Here is what to do, in order: read it once and note the signs (sudden jump, no class detail, a fake quote). Run it through your school’s detector if you have one, and treat the score as one weak signal. Then talk to the student: “Walk me through how you wrote this. Show me your notes. Explain this paragraph in your own words.” If they cannot explain the argument or show any working, follow your school’s process — and set the redo in class, with their notes. Then redesign the next essay so it asks for something from your classroom that AI cannot know.
The same four steps work for any subject — the signs, the detector’s limits, the conversation, and the redesign.
Common mistakes to avoid
- Accusing on a detector score alone. You will eventually accuse the wrong student — most likely a student writing in their second language.
- Ignoring the possibility that the student just wrote well. A good day is not proof of cheating. Ask before you conclude.
- Punishing without a conversation. A punishment without a conversation teaches nothing, and it destroys trust with the whole class.
- Banning AI outright. A ban does not stop AI use — it drives it underground, where you cannot see it or teach about it. Teach honest use instead.
- Setting homework AI can do in ten seconds. If the task is “write 500 words about the French Revolution”, AI will write 500 words about the French Revolution. The task is the problem.
FAQ
Can Turnitin detect AI writing?
Turnitin flags text it thinks is AI-written, but it has false positives and false negatives like every other detector. Treat a flag as a starting point for a conversation, not as proof.
Are non-native English speakers flagged more often?
Yes — detectors flag writing by non-native speakers at much higher rates. Be extra careful with these students, and never rely on a detector score alone.
Should I tell students I use detectors?
Be transparent about the process, not about specific tools. Tell students how work is checked and what counts as honest use. Secrecy does not prevent cheating; it prevents learning.
Can AI detectors be trusted at all?
As one signal among many, yes. As the only evidence, no. The conversation with the student, the drafts and notes, and your knowledge of the student are worth far more than any score.
