How Teachers Detect AI in Take-Home Assessments

The take-home assessment is, by design, the format where a teacher has the least control over how the work was produced. There is no exam hall, no invigilator, and no way of knowing whether a student spent four careful evenings on a piece or pasted a prompt into a chatbot at eleven o'clock the night before it was due. The whole point of a take-home task - depth, research, redrafting, working at one's own pace - is also exactly what makes it the easiest place to lean on a generative tool unnoticed. Learning to detect AI in take-home assessments is less about catching a single tell and more about building a fair, repeatable process for work that was, by definition, produced out of your sight. This guide is for teachers setting and marking unsupervised written tasks across GCSE, A-Level, and KS3.

Why Take-Home Tasks Are the Hardest Case

In a timed, supervised assessment you at least know the conditions: the student was in the room, alone with the paper, for a fixed window. A take-home task removes all of those guarantees at once. The student has unlimited time, full internet access, no supervision, and the privacy to use whatever tools they like. None of that is misconduct on its own - using a dictionary, a textbook, or a search engine is usually expected. The difficulty is that the same conditions that make legitimate research possible also make wholesale AI generation trivial, and the finished document looks identical either way.

This is why a take-home assessment cannot be policed the way an exam can. There is no controlled environment to rely on, so the burden shifts entirely onto the work itself and onto what you know about the student who supposedly produced it. The job is not to prove how the file was made - you usually cannot - but to notice when a piece does not fit the person, and to have a fair way of looking more closely.

What Does Not Fit the Student

The single most reliable signal in unsupervised work is a mismatch between the submission and everything else you have seen from that student. You mark these students all year. You know roughly how they write under timed conditions, how they argue in class, what their typical range is on a given topic. A take-home piece that suddenly reads two grades above a student's supervised work - fluent where they are usually halting, comprehensively structured where they usually drift, confidently using terminology they have never used in a lesson - is the thing to notice first. It is not proof of anything, but it is the strongest prompt to look harder.

Look, too, for the absence of the student's own fingerprints. Real take-home work usually carries traces of the person: a slightly idiosyncratic argument, an example drawn from something discussed in class, a misunderstanding carried over from earlier work, a voice you recognise. Generated work tends to be impersonal and generic - correct, complete, and oddly anonymous. The gap between a polished but faceless document and the specific student whose name is on it is where unsupervised AI use most often shows.

Reading the Likelihood Score

Your knowledge of the student tells you about fit; an AI detection tool tells you about the prose. The two together are far stronger than either alone, and for take-home work - where you have no conditions to fall back on - that second input matters more than usual. GradeOrbit's AI detection, included on school accounts, returns a score from 0 to 100 that measures how far the take-home piece departs from that student's own earlier work already on file - low means consistent with their own writing, high means it differs in ways worth a closer look - alongside a Low, Medium or High confidence label, the signals it noticed, and a short plain-English explanation. A high score on a take-home piece that already reads well above a student's supervised level is a clear prompt to look closer; a modest score on work that matches what you would expect from that student is reassurance.

As we explain in our guide on how AI detection likelihood scores work, the number is one input into a picture, never a verdict on its own - and that caution is doubly important for take-home tasks, where legitimate use of research and editing tools can nudge a score upward. GradeOrbit's AI detection is unlimited and included in every school teacher seat, whether you're triaging a class set of unsupervised submissions or looking closely at a take-home task that feeds into a reported grade. It works by comparing the submission against that student's own earlier writing on file, which carries more weight for a piece like this than a generic score would.

The Supervised-Work Comparison

Take-home assessments have one powerful advantage that fully unsupervised coursework often lacks: you almost always hold a supervised sample of the same student's writing. A timed essay, an in-class test, a mock script - any piece written under conditions you controlled gives you a baseline of how this student actually writes when they cannot use a tool. Setting the take-home submission next to that supervised sample is the most concrete check available. A large, unexplained jump in fluency, vocabulary, and structure between the two is far more telling than any single feature of the take-home piece on its own.

This is also a strong argument for building at least one supervised checkpoint into any course that relies on take-home work. A short, timed, in-class piece early in the year is not just an assessment in its own right - it is the reference point that makes every later unsupervised submission interpretable.

Turning Concern Into a Fair Conversation

When the fit, the supervised comparison, and the likelihood score all point the same way, the next step is a conversation, not an accusation. A take-home task gives you a natural opening because the student had time to engage deeply: "Talk me through how you approached this" is a fair, reasonable question. A student who genuinely did the work can describe their research, the choices they made, why they structured it as they did, and what they would change. A student who generated it will struggle to account for a document they did not really write.

Keep it curious rather than confrontational. Ask about a specific claim, a source they cite, a paragraph that sits oddly above their usual level. The aim is to move from your impression of the writing to factual questions the student can answer if the work is theirs. The wider framework sits in how teachers handle a high AI likelihood score fairly, and for the difficult middle cases the guide above is worth reading alongside it. Detection should never replace your professional judgment; on an unsupervised task especially, a high score is a reason to ask, not a conviction.

The EAL and Careful-Writer Caveat

Unsupervised conditions give every student time to edit, and the students who use that time most thoroughly - careful writers, conscientious EAL students, anyone who drafts and polishes - can produce prose that scores higher than their actual authorship warrants. A student who legitimately spent four evenings refining a take-home essay has, in effect, removed exactly the roughness that distinguishes human writing from generated text. This is precisely why the supervised comparison matters so much for take-home work: it gives you evidence of the student's authorship that is independent of how polished the final piece is. A student whose take-home work is an entirely plausible step up from their supervised baseline has done the work, whatever the prose score suggests.

Documenting the Check for Moderation

Whatever you conclude, record it at the time. For a take-home assessment, a clear note covers what you compared and what you found: "Take-home essay likelihood 81%; reads roughly two grades above student's December mock written under timed conditions; student unable to explain the central argument when asked." That is a factual, defensible record that holds up if the case is revisited later. Where take-home work feeds into internally assessed components, keep the detection run and your comparison notes alongside your marking - the guidance above covers how this fits the wider evidence pack. If your school's policy does not yet address unsupervised AI use specifically, the school AI academic integrity policy guide is a sensible place to start.

Running Take-Home Work Through GradeOrbit

AI detection in GradeOrbit runs on school accounts and follows the same process for every subject. You run it from a student's profile, or on a whole class at once. For handwritten take-home work, upload scans or photos, or use the phone QR camera link; typed submissions go in as a PDF or image (export Word documents to PDF first). For a class set, bulk upload splits the pages by pages per student and reads the name on each cover page to match it to your class list, asking you to pick the student if the match isn't clear. Each piece is compared with that student's own earlier work on file - a student with no earlier work is skipped, and you are told - and results appear on the student's profile. Pupils' work is held under the school's Data Processing Agreement, scans are removed when the run archives, and nothing is used to train AI models. The record of the detection, your supervised comparison, and your decision belongs in your own notes and your school's systems.

Try GradeOrbit for AI Detection in Take-Home Work

If you want a detection check that compares each take-home submission with the student's own earlier work, to sit alongside your knowledge of the student and your professional judgment, GradeOrbit's school accounts include it for every teacher, with no per-check charge. It isn't offered on individual teacher accounts, because a generic "does this read like AI?" score on a single piece isn't reliable enough to raise with a student.

See how AI detection works for schools and start a free school trial for your department.