How Teachers Handle AI Scores on Year 13 EPQ Submissions

The EPQ is different from almost everything else you mark. By the time a Year 13 student hands you their final 5,000-word report, they have lived with it for the better part of a year, you have signed off their production log at three or four checkpoints, and the submission deadline to the exam board is days away. So when you run that final report through a detection tool and the AI likelihood score comes back high, the stakes feel very different from a routine homework flag. There is no "redraft it and resubmit" - the work is about to leave your hands and go to a moderator.

For a supervisor or coordinator looking at a high score on a final EPQ submission, the question is not "what signals reveal AI use?" - it is "the score came back high, the deadline is close, and I need to do this fairly. What now?" The honest answer is that the process is the same fair process you would run on any flagged work, but the EPQ gives you more evidence to work with and a tighter clock to work within.

Why Final-Submission Scores Feel Higher-Stakes

A flagged piece of homework can be re-set. A flagged GCSE mock can be discounted. A final EPQ report heading to the exam board cannot be quietly redone - it is a formal qualification submission, and signing the candidate declaration on work you have doubts about puts your professional integrity on the line, not just the student's. That is why the instinct to either rush to accusation or quietly ignore the score is so strong, and why both are wrong.

It helps to remember what the number actually is. Most detection tools return a probabilistic score between 0% and 100% that measures how closely the linguistic patterns in the text resemble patterns commonly produced by large language models - sentence rhythm, vocabulary distribution, predictability of phrasing. GradeOrbit's school detection asks a narrower question: its 0 to 100 score measures how far the report departs from the student's own earlier writing. Either way, a 90 is not a confession and not a verdict. It is one piece of statistical evidence, and on the EPQ you happen to have an unusually rich pile of other evidence to weigh it against.

The Evidence You Already Have

This is where the EPQ is genuinely easier than a one-off essay. You are not starting from a blank page - you have a documented trail built over months. Pull it together before you speak to anyone.

• The production log. The student has recorded their planning, research, and decision-making at each stage. Does the voice and thinking in the log match the voice in the final report? A report written in fluent academic prose attached to a thin, generic production log is more interesting than the score alone.
• Supervisor meeting notes. You met this student. You discussed their argument, their sources, their struggles. Does the finished report reflect the project you watched develop, or has it drifted into territory you never discussed?
• Earlier drafts. Most EPQs go through draft stages. Read the progression. A first draft at one level of polish and a final draft that suddenly reads like a different author wrote it is a signal worth examining.
• The viva or presentation. If your centre runs a presentation element, you have heard the student talk about their own work unscripted. That memory is evidence too.

Document what you find in a short, factual note: "94% likelihood on final report; production log voice inconsistent with report; second draft markedly less polished than final; presentation fluent." That contemporaneous record is what protects you if this escalates - and on a qualification submission, it might.

Running the Conversation Before the Deadline

The clock matters here, but it must not turn the conversation into an interrogation. Find time to speak to the student before the submission deadline, and open with the work, not the score. Ask them to walk you through the hardest part of the project, why they framed their research question the way they did, which source changed their thinking, what they would do differently with another month. A student who genuinely wrote the report can usually talk about it fluently - the thinking is in their head. A student who generated large parts of it tends to repeat surface points and struggle when pressed on a specific choice.

If you raise the detection result, frame it as information rather than proof. "An automated check flagged parts of this report as similar to AI-generated writing, and the polish reads differently from your earlier drafts. I wanted to talk to you about it" invites a response. "The AI detector says you cheated" invites a defensive shutdown and teaches you nothing. The same fair-process principles apply here as in our guide on handling a high AI likelihood score fairly - the EPQ just compresses the timeline.

Listen for the boring middle. Most cases are not "wrote it all themselves" or "generated the whole thing." They are "used a chatbot to structure the literature review, then wrote the analysis themselves" or "drafted it independently, then asked AI to tighten the prose." Your centre's policy and the exam board's malpractice rules treat those middle cases differently from wholesale generation - but only if you find out which one you are dealing with before you sign the declaration.

Aligning With the EPQ Mark Scheme and Malpractice Rules

The detection score does not decide the outcome - your exam board's malpractice procedures and your centre's academic integrity policy do. Read both before the conversation, not after. EPQ specifications require the work to be the candidate's own and require you, as supervisor, to authenticate it. If your conversation and evidence leave you genuinely unable to authenticate substantial sections, that is a malpractice matter that goes to your exams officer and, through them, potentially to the awarding body - on their timeline and their forms, not yours.

Conversely, if the student demonstrates clear ownership and the score turns out to reflect a polished but genuine writer, you authenticate the work and move on. EAL students, students who have spent a year reading academic prose for the project, and students who write in a deliberately formal register all produce text that detection models flag as AI-like. The score being high does not override what the conversation and the production log tell you. Log the detection result in the moderation pack regardless - a moderator may want that context, and the moderation cycle is exactly where a score earns its place as one input among many.

How GradeOrbit Supports the Process

On a GradeOrbit school account, unlimited AI detection is included in every teacher seat, whether it's a lower-stakes draft or a final submission heading to the exam board - there is no per-check charge and no separate tier to reach for. Under your school's signed Data Processing Agreement (DPA), it works by comparing the report against work that student has already had marked in GradeOrbit, which is a stronger basis for a final EPQ authentication than a generic score would be. You can run it from the student's profile, upload the report as a PDF or scanned pages, and the result - a score, a Low, Medium or High confidence label, the signals it noticed and a plain-English explanation - appears on their profile. A student with no earlier work on file is skipped, and you are told. The score is still the start of the fair process above, never a substitute for it.

How student data is handled depends on your account type: on a school account, students' work is held under the DPA with their real names so each check can compare against that learner's earlier writing on file. Uploaded scans are removed when a run archives (90 days by default, adjustable by the school between 7 and 365 days), and student 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, students are referred to only as "Student 1", "Student 2", and work is never saved to a database (processed and discarded immediately). Across all tiers, student work is never used to train AI models. The audit trail that matters is in your production logs, your meeting notes, and your markbook. For the language to use in the conversation itself, our guide on the student conversation goes deeper.

The Supervisor's Call

On a final EPQ, a detection score sits next to the production log, the drafts and the conversation as one more piece of evidence. It does not authenticate the work and it does not condemn it. That decision is yours as supervisor, and it is the decision the exam board relies on when you sign the declaration.

GradeOrbit for Schools includes detection in every teacher seat, and a school can start with a free school trial so a team can test detection and marking on real work. The data-processing terms are set out at gradeorbit.co.uk/schools.