How to Detect AI in A-Level Geography NEA Reports

The A-Level Geography NEA - the independent investigation - is one of the few coursework components built almost entirely on a student's own primary data. They choose a question, collect their own fieldwork, run their own analysis and write it up. That structure is exactly why learning to detect AI in NEA reports fairly matters: the parts of the report that should be tied to a specific river transect, beach survey or urban land-use count are the parts a generative tool cannot genuinely produce, even when it writes around them convincingly.

A common moment: you are working through a stack of 3,000-to-4,000-word investigations against the deadline, and you reach a literature review or evaluation that reads more fluently than the rest of the report. The honest position is that you cannot prove authorship from the text alone - but you can run a fair, repeatable check that turns a suspicion into a defensible professional judgment rather than a hunch.

Where AI Slips Into an NEA Report

An NEA report has a predictable skeleton: an introduction and question, a literature and theory review, a methodology, data presentation, analysis, a conclusion and an evaluation. The risk is not spread evenly across those sections. The data presentation and analysis are anchored to the student's own numbers and maps, so AI-generated prose there tends to drift away from the actual figures. The literature review, the theoretical context and the evaluation, by contrast, are general enough that a chatbot can write a smooth, plausible version with no fieldwork at all.

That uneven risk is useful. A report where the analysis is hesitant and specific but the literature review is suddenly confident, broad and stylistically different is showing you exactly the seam worth examining. The mismatch between a student wrestling with their own messy data and a polished, generic discussion of "the Burgess concentric zone model" or "fluvial processes in general" is often more telling than any single sentence.

Reading a Likelihood Score on an NEA Section

The first habit to build is the right mental model for what an AI detection tool gives you. A likelihood score is a signal, not a verdict. GradeOrbit's detection tool returns a probability between 0% and 100% that the text under review was AI-generated, along with a short, plain-English explanation. It does not return a finding, and it certainly does not return a confession. It tells you where to look more closely.

Treat the number like a temperature reading rather than a result. A methodology section at 11% likelihood needs no second glance unless something else has already caught your eye. An evaluation at 86% does not warrant an accusation - it warrants a closer read, a comparison against the student's known work, and, where your process requires it, a second marker. Because an NEA is long, run sections separately rather than pasting the whole report in at once: a flagged literature review tells you something a whole-document average would hide. On a GradeOrbit school account, unlimited AI detection is included in every teacher seat with no per-check credit charge, and under your school's Data Processing Agreement (DPA) it compares the section against that student's own earlier writing already on file, which is a more useful signal for a close call than a generic one-off score would be. Our wider guidance on how to interpret AI detection likelihood scores goes deeper on reading these numbers.

Cross-Check Against the Fieldwork

The single strongest move with an NEA is something no generic coursework has: you can put the writing next to the data it is supposed to describe. A genuine analysis section references the student's own results in specific, sometimes awkward ways - "my Spearman's rank value of 0.62 was lower than expected, probably because site three was disrupted by roadworks on the day". AI-generated analysis tends to be confidently generic, discussing what such a study "would typically find" rather than what this student actually measured.

Read the suspect section with the data tables, field sketches and raw recording sheets open beside it. Do the named figures match? Does the evaluation reflect the real limitations of the day - the broken equipment, the rained-off second visit, the sample that was smaller than planned - or does it list textbook limitations that could attach to any study? A report that is fluent but strangely disconnected from its own fieldwork is showing you a gap worth following up, and that gap is harder to explain away than style alone. For a related subject lens, our note on how to detect AI in GCSE geography coursework covers the same instinct at Key Stage 4.

How to Handle a High Score Fairly

When a section returns a high likelihood score, compare it against writing you already know is the student's. NEAs are well suited to this because the work leaves a long paper trail: the proposal, the fieldwork plan, lesson-time progress checks, handwritten field notes and earlier drafts. Compare the rhythm, the vocabulary range and the kinds of mistakes the student usually makes. Genuine student writing has a fingerprint that generative AI rarely reproduces.

If, after that comparison, you are still unsure, log the case as ambiguous and follow your school's process rather than forcing a conclusion. A detection score should never be the sole basis for a decision that affects a grade or a student's record - and an NEA carries a moderation and appeal trail, so your reasoning needs to be the kind of thing you would be comfortable explaining to a Head of Department or an exam board. Our walkthrough on how teachers handle a high AI likelihood score fairly sets out that conversation step by step.

Keeping the Process Fair and Documented

Consistency is what makes detection defensible: the same threshold for when a section gets a closer look, the same next step when one is flagged, and a short written note of the score, your professional read and the decision made. The goal is not a file of accusations - it is a record that shows you applied judgment fairly and the same way for every student.

It also means protecting student data while you do it. How student data is handled depends on your account type: because teacher AI detection compares an NEA report against earlier writing on file, it runs on GradeOrbit school accounts under a signed Data Processing Agreement (DPA), where student work is retained securely for your school's agreed retention window (three years by default, and wiped automatically on schedule or offboarding). By contrast, on an individual teacher account where no school DPA is in place, you redact personal details client-side in your browser before marking, the work is never stored, and the file is gone the moment marking is done. Across all tiers, student work is never used to train AI models.

In Short

Treat the NEA the way you already treat moderation: calmly and on evidence. A likelihood score is a prompt to look again, the student's own fieldwork is the cross-check, and a documented decision is the outcome. Detection informs your professional judgment; it does not decide anything on its own.

If your department wants a baseline-aware detection step for NEA marking, the school plan is described at gradeorbit.co.uk/schools, and departments can trial it for 30 days.