EPQ AI Checker: How to Detect AI in an A-Level EPQ Fairly

An Extended Project Qualification asks a student to research independently for months, keep a production log, and produce a 5,000-word essay or a related practical artefact. Most of that work happens outside lessons, with a supervisor who sees it in snapshots. That is exactly the setting in which undeclared use of ChatGPT or Claude is hardest to notice.

For supervisors, the tension is obvious. You are responsible for authenticating the work, but telling a student you think their EPQ was written by AI is a serious allegation that can damage their confidence and their progress. Plagiarism checkers were not built for this problem, and a likelihood score only helps if it is read alongside what you know of the student. If you are still choosing a tool, our guide to AI detection tools for teachers covers what actually matters for classroom use.

The Challenges of Spotting AI in an A-Level EPQ

Spotting AI-generated content in a 5,000-word project is a different job from catching a paragraph pasted from Wikipedia into a homework task. Generative AI models do not simply copy existing text; they predict the next most likely word based on vast training datasets, creating entirely novel sentences that will completely bypass traditional plagiarism checkers. This means that a student could theoretically generate an entire EPQ literature review that scores a perfect 0% on standard similarity software.

Why Traditional Plagiarism Checkers Fail

Traditional plagiarism detection software works through exact string matching and structural similarity comparisons against an established database of published works, past essays, and internet pages. When a student uses Claude or ChatGPT to write their EPQ, the resulting text is technically original. The AI synthesises the information, so a similarity checker is searching for a source that does not exist. A clean similarity report tells you the work was not copied; it tells you nothing about whether it was generated.

The Risk of False Positives

The converse problem is just as dangerous: the risk of false positives. Many teachers have turned to early AI detection tools that operate purely by looking for 'burstiness' (variation in sentence length) and 'perplexity' (how predictable the vocabulary is). Unfortunately, these metrics often penalise students who write in a highly structured, academic, and formulaic manner - exactly the kind of writing style that the A-Level EPQ encourages. Falsely accusing a high-achieving student of using AI because they wrote a clear, methodical essay can destroy the fragile teacher-student trust and demotivate the student entirely.

Using Likelihood Scores and Probabilities

Detection tools have largely moved away from binary "AI vs. Human" verdicts towards probabilities. When reviewing an EPQ, you are not looking for certainty; you are looking for an indication of how likely it is that AI was heavily involved in the drafting.

Moving Away From Absolute Certainty

No software in the world can currently guarantee with 100% accuracy whether a specific paragraph was written by a human or a machine. Generative models are constantly evolving, and human writing styles are incredibly diverse. Any check on an EPQ comes with a margin of error. Treating a score as "a reason to look further" rather than "proof of cheating" is the single most useful change a department can make to its academic integrity policy.

Reading a Score Against the Student's Own Writing

GradeOrbit's AI detection, included on school accounts, returns a score from 0 to 100 that measures how far the new work departs from that student's own earlier work already on file, with a Low, Medium or High confidence label, the signals it noticed, and a plain-English explanation. A score of 85 does not mean 85% of the EPQ was written by AI. It means the project reads very differently from how this student has written before, in ways worth a closer look. A low score means you can carry on with your normal marking; a high one is the cue for the human evaluation described below, never a finding in itself.

That comparison has a practical consequence for the EPQ. A student with no earlier work on file is skipped rather than given a generic score, and you are told. If your EPQ cohort is drawn from several subjects, make sure each student has some genuine earlier writing in their profile, such as a marked essay from one of their A-Level subjects, before the final product comes in, or the check has nothing to compare against.

Combining AI Detection with Professional Judgment

Software should never decide a student's outcome on its own. A high likelihood score on a section of an EPQ needs to be read alongside what you know as the supervisor: the student's usual standard, and how the project has developed over the year. None of that is visible to a detection tool.

Cross-Referencing the EPQ Logbook

The EPQ has a built-in safeguard: the production log. This logbook is supposed to comprehensively document the student's journey, detailing their planning, research struggles, shifts in project focus, and drafting timelines. If a student submits a flawless 5,000-word essay with a high AI likelihood score, but their production log is sparse, generic, or lacks specific timeline details, this provides strong circumstantial evidence of AI generation. Conversely, if a student has an extensively detailed logbook that perfectly aligns with a highly sophisticated final essay, a high AI score might merely indicate the use of AI for proofreading or restructuring, which you can discuss with the student.

Recognising the Student's Authentic Voice

By sixth form, most students have a recognisable academic voice. While the EPQ expects a step up in academic terminology, drastic and sudden shifts in tone, vocabulary, and syntactic complexity are major red flags. If a student who typically struggles with basic comma splices is suddenly producing perfectly balanced, multisyllabic arguments about complex socioeconomic theories, your professional judgment should immediately be engaged. You can cross-reference the EPQ with handwritten mock exams or previous controlled assessments to compare the genuine voice with the submitted text.

Having the Conversation with Students

If the score is high and your own reading also suggests the project is not entirely the student's work, approach the conversation with care. Inquisitorial tactics often backfire, leading to defensive denials. Instead, use an investigative, supportive approach.

Constructive Approaches to Suspected AI Usage

Rather than demanding a confession, ask the student to explain specific, complex concepts from their EPQ. Ask them to justify why they chose a particular source over another, or how they synthesised conflicting viewpoints in their conclusion. If a student has genuinely researched and written the piece, they will be able to discuss the topic fluently, even if they stumble over their words. If they used ChatGPT to generate the entire essay, they will quickly find themselves unable to explain the nuances of their own submitted work. The conversation is also a chance to explain the difference between using AI to brainstorm and using it to do the thinking the project is meant to assess.

What JCQ Expects Before You Sign the Declaration

"Can you use AI for an EPQ?" is one of the questions students search for most, and the answer comes from JCQ's guidance on AI use in assessments, which covers the EPQ as it does other non-exam assessment. Students may only use AI tools where the centre allows it, they must acknowledge any use they make of them, and the work submitted for assessment must be their own. Using AI to explore a topic or find sources the student then reads is very different from submitting text it wrote, and the production log is the natural place for students to record what they used and why.

Timing matters. If you have concerns before the student signs the declaration of authentication, you deal with them within the centre: talk to the student, and where appropriate have them redo the affected section in their own words. Once the declaration is signed and the work is submitted, suspected misuse is a potential malpractice case that goes to the awarding body under your centre's procedures. That is the strongest argument for checking drafts at the milestones the log already records, rather than waiting for the final product.

Marking the EPQ Itself

Many people who search for an "EPQ marker" or "EPQ grader" are really asking who marks the project. The EPQ is centre-marked: the supervisor marks against the awarding body's assessment criteria, the centre standardises internally, and the board moderates a sample. On AQA's specification, the 50 marks are split across four objectives: AO1 Manage (10), AO2 Use resources (10), AO3 Develop and realise (20) and AO4 Review (10). Much of AO1 and AO4 is evidenced by the production log and the presentation rather than the written report, which is one more reason no tool can mark an EPQ on its own.

Where AI can help is as a second read of the written report before internal standardisation. On an individual teacher account, export the final report to PDF, paste the assessment criteria as the mark scheme, and GradeOrbit returns a suggested mark and feedback against those criteria for 1 credit per student script, and a report of EPQ length sits comfortably within the page limit. The work is processed and discarded rather than stored. Treat the output as a check on your own reading of AO2 and AO3 evidence, never as the mark you submit. If you are weighing up tools for this, we compare the best AI marking tools on how well each one handles a real UK mark scheme.

In Short

Similarity checkers will not catch generated text, a likelihood score is a probability rather than a finding, and the production log plus a conversation with the student will usually tell you more than either. The aim is not to police every keystroke but to be confident that the project you sign off is the student's own.

If your centre uses GradeOrbit for Schools, detection is included in every school teacher seat and each check compares the submission against that student's own earlier writing on file, under your school's Data Processing Agreement. There is no depth tier to choose first, and school accounts start with a 30-day trial.

Further reading: detecting AI in A-Level English coursework and how do teachers check for AI?