How Teachers Detect AI in A-Level Computer Science NEA

If you want to detect AI in A-Level Computer Science NEA work, the first thing to be clear about is where the risk actually sits. The programming artefact itself — the code a student writes and demonstrates — is only part of the qualification. The marks-heavy part is the written report: the analysis of the problem, the design decisions, the technical solution walkthrough, the testing evidence and the evaluation. That is thousands of words of formal, structured prose, produced over months at home, and it is exactly the kind of writing that tools like ChatGPT and Claude can generate convincingly. This guide is for UK A-Level computing teachers who want to check that writing fairly, without turning a suspicion into an accusation on the strength of a single number.

The NEA is a big investment for a student and a big marking job for you, so getting the integrity question right matters on both sides. The aim is not to police, but to be confident that the report you are crediting reflects the candidate's own understanding of their own project.

Why the NEA Report Is a Soft Target for AI

Coursework written away from the classroom over an extended period is the highest-risk category for AI use, and the Computer Science NEA has several features that make it particularly tempting to outsource. The report follows a predictable structure that a language model handles well: a stakeholder analysis, success criteria, a design rationale, an account of the algorithms used, and a reflective evaluation against the original objectives.

• Formulaic sections. "Analysis" and "Evaluation" have conventional shapes that AI reproduces fluently.
• Technical vocabulary as cover. A dense paragraph about time complexity or normalisation can read as competence when it is actually generic filler.
• A gap between code and commentary. A student can write real, working code and then have AI narrate it — so the artefact is genuine but the marked write-up is not.
• Volume. The sheer length of the report makes a careful human read slow, which is where a first-pass signal helps.

That last point is the crux. You are not looking for a plagiarised passage you can match to a source; you are trying to judge whether the analysis and evaluation genuinely came from the student who built the project.

Detection Is Probabilistic, Not a Verdict

The single most important thing to understand about any AI detection tool is that it produces a likelihood, not a fact. GradeOrbit's built-in AI detection gives each submission a score from 0 to 100% representing how likely the text is to be AI-generated. That score is a prompt for your professional judgment — never a substitute for it. A high number tells you where to look closely; it does not, on its own, prove misconduct, and no responsible process should treat it as if it did.

This matters especially for computer science, where a candidate's authentic voice can be sparse and technical. A student who writes in short, precise, textbook-flavoured sentences may sit higher on a likelihood scale than a chatty essayist, without having done anything wrong. Reading the score alongside what you already know about the student, their earlier drafts, and their classwork is the whole discipline. For a fuller account of the reasoning behind these numbers, our guide on how AI detection likelihood scores work is a good companion.

Reading the Score Section by Section

A blanket score for a whole NEA report is less useful than a sense of which parts sit high. Because AI use is often patchy — a student who wrote their own design section but generated the evaluation the night before the deadline — it helps to look at the report in its natural parts rather than as one block. Run the analysis, the design rationale and the evaluation separately where you can, and see whether the signal clusters in the reflective, prose-heavy sections that are easiest to fake and hardest for you to verify against the code.

GradeOrbit is designed to sit inside your normal marking, so the same tool you use to work through the report can flag the likelihood as you go. That keeps detection as one input among several rather than a separate, adversarial exercise bolted on at the end. If a section reads as high-likelihood, the productive next step is a conversation, not a charge — which our piece on how to talk to students about AI detection results walks through.

Triangulating Against the Project Itself

Computer Science gives you a verification advantage that essay subjects do not: the report is supposed to describe a real, running artefact. That means you can cross-check the writing against the code and against the student's ability to explain it. A viva-style chat — "walk me through why you chose this data structure", "show me the test that catches this edge case" — is often more decisive than any score. If the report confidently discusses an algorithm the student cannot explain at their own machine, the likelihood signal has done its job by pointing you there.

Keep a light record of what you checked and why. If you are ever asked to justify a decision to a moderator, a head of department, or a parent, a short note showing that you used the score as one signal, looked at drafts, and held a professional conversation is far stronger evidence than a screenshot of a percentage. Our guide on how teachers document AI detection decisions for records covers a proportionate way to do this.

Keeping It Fair for Every Candidate

Fairness cuts both ways. A student who worked genuinely hard on a strong report deserves not to be under a cloud because a tool returned a middling number, and a student who did outsource their evaluation deserves to be asked about it rather than quietly downgraded. The exam boards and JCQ are clear that misconduct decisions rest on evidence and process, not on an automated output. Treat the detection score as the start of an enquiry, apply the same standard to the whole cohort, and let your knowledge of the candidate carry the final call. Used this way, detection supports academic integrity without punishing an honest, plainly-written student for the crime of sounding tidy.

Try GradeOrbit for NEA Integrity Checks

GradeOrbit's AI detection gives you a 0-100% likelihood score you can run across the written sections of an A-Level Computer Science NEA, so you know where to focus a closer read rather than working blind through thousands of words. It is assistive by design: the tool surfaces a signal, and you — with the code in front of you and the student's understanding to test — make the judgment. That is how detection should work, as a prompt for your expertise and not a replacement for it.

Sign up to GradeOrbit and try the built-in AI detection on your next set of NEA reports.