How Teachers Detect AI in Online Homework Submissions
Homework set and collected online - through Google Classroom, Microsoft Teams, Show My Homework, or a simple shared document - has quietly become the single most exposed piece of work a teacher hands out. It is done unsupervised, at home, on the same device a student uses to access an AI chatbot, and it is submitted as typed text that can be generated in seconds. Teachers who are perfectly comfortable judging a piece of classwork suddenly find themselves staring at a fluent paragraph and wondering whether the student wrote a word of it. Checking that work with a fair, repeatable process rather than a hunch is now part of the everyday job, and the detection score is where that inquiry starts, not where it ends.
Why Online Homework Is So Exposed
The conditions around online homework line up almost perfectly for AI use. The task is set remotely, completed without supervision, and returned as editable text rather than a photographed page. The student is already on a device, the prompt - "write a paragraph explaining the causes of the First World War" - is exactly the kind of self-contained question a large language model answers comfortably, and the temptation to paste a polished answer and move on to the rest of the evening is real. None of that makes a class of cheats; most homework is still genuinely the student's own. But because the work is produced in exactly the circumstances that make AI use easiest, a routine, transparent check protects the students who did the work and gives you a fair basis for a conversation with those who may not have.
It is worth being honest about the scale of the problem too. A great deal of AI use on homework is not calculated dishonesty; it is a tired or anxious student reaching for help with a task they find hard, late in the day. That framing matters because it shapes how you respond. The goal of detection here is not to build a case against a fourteen-year-old. It is to keep homework meaningful - to make sure the practice and the thinking the task was set to produce actually happened - and to teach students why doing the work themselves matters before the stakes get higher.
What GradeOrbit's Likelihood Score Tells You
GradeOrbit's AI detection, which is part of its school accounts, returns a score from 0 to 100 that measures how far a new piece departs from that pupil's own earlier writing on file. A low score means the homework is consistent with how they usually write; a high score means it differs in ways worth a closer look. Alongside the number you get a Low, Medium or High confidence label, the signals it noticed, and a short plain-English explanation. It does not tell you with certainty that a particular student used AI, and it should never be read as a yes-or-no answer. It is an indicator to review, not proof.
For homework specifically, two things are worth holding in mind. First, homework tasks are often short, and the shorter the sample the more cautiously the number should be read - a single paragraph carries less signal than a full essay, so a raised score on a brief task is a flag to look closer, not proof. Second, online homework frequently follows a structure the class has been taught explicitly, which can nudge genuine answers towards a similar, slightly formulaic register. A capable student following the method they were shown can share surface features with AI output. The score earns its keep by telling you which submissions to read more carefully, not by deciding the question for you. Our guide on how AI detection likelihood scores work goes into more detail on what the number does and does not mean.
On a GradeOrbit school account, unlimited AI detection is included in every teacher seat, whether you're running a quick check across a class set of online submissions or looking closely at the one piece you intend to discuss with a student, with no per-check charge and no credits to manage. Under your school's Data Processing Agreement (DPA), it compares the submission against work that pupil has already had marked in GradeOrbit, so the comparison is with their own writing rather than a generic idea of what AI sounds like.
Reading a High Score Fairly
A high score on a homework submission warrants a closer look, never an accusation. Your most valuable reference point is the student's own supervised writing - classwork, exit tickets, anything produced in front of you. Read the flagged submission alongside it. Does it sound like the same young person? Is the vocabulary, sentence shape, and depth consistent with what they produce in the room, or does the homework suddenly acquire a polish and fluency their classwork never shows? A submission that is markedly more sophisticated than everything the student has written under your eye is the pattern worth exploring.
Genuine homework carries the texture of a real student working at home: the occasional clumsy sentence, the idea half-expressed, the example drawn from a lesson they actually sat in. AI-generated answers tend to be smoothly competent but generic - well balanced, confidently structured, and oddly free of the specific, slightly uneven detail a particular student would bring. Treat the number as one input you weigh against your own knowledge of the student, never a substitute for it.
Turning Detection Into a Conversation, Not an Accusation
If both the score and your own read raise concerns, the next step is a low-key conversation built around the content, not the detection result. Ask the student to talk you through their answer: "Explain this point to me in your own words," or "Why did you choose this example?" A student who did the thinking can expand on it naturally; a student who generated it usually cannot add the specific reasoning that genuine work produces. The aim is to find out what happened and to put the homework right, not to extract a confession.
Because so much homework AI use comes from pressure rather than dishonesty, the framing of that conversation does a lot of work. Make clear that the point of the homework was the practice, that copying an answer skips exactly the thinking the task was meant to build, and then help the student redo it properly. For younger students in particular, an early, supportive conversation about doing their own work is far more valuable than a sanction.
Keeping a Sensible Record
Most homework will never need a record beyond your markbook, but where you have raised a concern with a student a brief contemporaneous note is good practice. A couple of sentences will do: the score, the nature of the concern, what you discussed, and the outcome. If the redone work is clearly the student's own, that is the end of the matter and worth recording as such. If a pattern develops across several pieces, that note is what turns a series of hunches into something you can take to a head of department or raise with parents fairly.
Running Online Homework Through GradeOrbit
On a school account you can check one submission from the pupil's profile, or run a whole class at once. Typed homework goes in as a PDF (export it from Classroom, Teams or Word first, as Word files aren't accepted); photographed or scanned pages go in as images, or straight from your phone using the QR camera link. For a class set, bulk upload splits the pages by the number of pages per student and reads the name on each cover page to match it to your class list, asking you to pick the pupil when the match isn't clear. The check compares each piece with work that pupil has already had marked, so the more of their earlier work is on file, the more useful the result; a pupil with nothing on file yet is skipped and you are told, rather than given a generic score. Results appear on each pupil's profile.
How student data is handled depends on your account type: because teacher AI detection compares a submission against earlier writing on file, it runs on GradeOrbit school accounts under a signed Data Processing Agreement (DPA), where pupils' work is held with their real names. Uploaded scans are removed when a run archives (after 90 days by default, and the school can set anything from 7 to 365 days), and pupil records follow the school's retention window and are wiped at offboarding. By contrast, on an individual teacher account where no school DPA is in place, uploaded student work for marking is redacted client-side in your browser, labelled anonymously, and never stored (processed and discarded immediately). Across every tier, student work is never used to train AI models.
Keeping Homework Worth Setting
The reason to check online homework at all is to keep it meaningful, so the practice the task was set for actually happens. A likelihood score helps you decide which submissions to read more closely; your knowledge of the student decides what the score means.
Detection is part of GradeOrbit for Schools, and every school account can start with a free school trial, so it can be tried on real student writing first. The full details are at gradeorbit.co.uk/schools.