How to Use AI Detection Results as Professional Evidence

An AI detection score appears on your screen. It reads 79%. The confidence label says High. Now what? For many UK secondary teachers, this is the moment the tool hands the problem back to you - and the path forward is far from obvious. The number feels authoritative, but acting on it without a structured process puts both you and the student in a difficult position.

This guide is for teachers who want to use AI detection results as professional evidence: not as an automated verdict, but as one strand in a considered, documented, defensible case. Getting this right matters for academic integrity. It matters even more for the welfare of the students you are assessing.

What a Likelihood Score Actually Tells You

AI detection tools do not tell you whether a student used ChatGPT, Claude, or any other tool. They tell you how closely the submitted text resembles patterns statistically associated with AI-generated output. That distinction is not a technicality - it determines what you can and cannot conclude from the result.

GradeOrbit's AI detection, available on school accounts, takes a more specific approach: it compares the new work with that pupil's own earlier work already marked in GradeOrbit, and returns a score from 0 to 100 for how far the new work departs from it, alongside a confidence label (Low, Medium, or High), a list of the specific signals it noticed, and a plain-English reasoning paragraph. That reasoning - the signals themselves - is what makes the result professionally usable. A score alone, without explanation, is a number. A score with explained reasoning is evidence you can articulate to a colleague, a Head of Department, or, if it comes to it, a formal panel.

The Joint Council for Qualifications (JCQ) is clear that teachers bear the responsibility for authenticating student work, and that detection tool outputs should never be used as the sole basis for an academic integrity decision. A likelihood score is a starting point for professional inquiry - not a conclusion.

Class Screening vs. a Single Pupil - Running the Check

In GradeOrbit, detection is unlimited and included in every school teacher seat, so there is no per-check cost to weigh and no mode to choose. You can run it on a whole class at once, or from a single pupil's profile when one piece concerns you. Work goes in as scans, photos or PDFs, or through the phone camera link; for a class set, bulk upload splits the pages by pages per student and matches each cover-page name to your class list.

The practical guidance is straightforward. Use a whole-class run for initial screening - particularly when you are working through a class set and want to identify pieces that warrant closer attention. For higher-stakes work (coursework, NEA, assessed controlled assignments), or when you are already building a documented concern, spend your time on the reasoning paragraph and signals for that pupil, and check them against the earlier work yourself.

Because the comparison depends on earlier work, a pupil with no earlier work on file is skipped, and you are told, rather than given a generic score. That is deliberate: a one-off "does this read like AI" number is not reliable enough to raise with a pupil.

Combining the Score With Your Professional Knowledge

The most powerful tool in any academic integrity case is not the detection software - it is your knowledge of the student. Before taking any action on a high score, ask yourself a structured set of questions.

Does this piece of work feel consistent with what this student normally produces? A student who typically writes at Grade 4 and submits work that reads at Grade 7 is a different situation from a high-achieving student whose detection score is elevated. If you have previous examples - class exercises, timed tasks, rough drafts - compare them directly. An unexplained jump in quality, fluency, or structural sophistication is meaningful context regardless of the score.

Is the register of the writing consistent with the subject? Formal academic subjects - History, Religious Studies, Sociology, Law - require writing that can appear unusually polished or structured. Strong students who have internalised the subject register may produce work that detection tools misread as AI-generated precisely because they have mastered the conventions. The same applies to EAL students who draft in their first language before translating: the resulting English can have a slightly over-structured quality that scores poorly without any AI involvement.

Were there any contextual signals around submission? Work submitted late at night immediately before a deadline, from a student who typically struggles, represents a different risk profile from the same score on a piece submitted days early by a consistently strong writer.

Documenting a Concern Properly

If your professional assessment - score plus linguistic signals plus your knowledge of the student - points to a genuine concern, the next step is documentation before any conversation with the student. Write down the following, clearly and contemporaneously:

• The detection score, confidence label, and the specific signals the tool identified
• Your own assessment of the text - what in the writing itself contributed to your concern
• A brief summary of the student's prior work and whether this piece is consistent with it
• Any contextual factors (submission timing, access to devices, previous conversations)

This record serves two purposes. It protects you by demonstrating that your concern is grounded in professional judgment rather than a single automated score. And it protects the student by ensuring that any subsequent conversation or formal process is based on specifics they can respond to, not a number they cannot challenge.

Having the Conversation With the Student

If you decide to raise the concern directly, the conversation should be exploratory rather than accusatory. Students who wrote their own work can discuss it. Students who submitted AI-generated content often cannot.

Ask open questions about their argument: what was their main claim, where did they find a particular piece of evidence, what counterarguments did they consider? If the work involved a structured response to source material, ask them to explain their interpretation in their own words. Ask them to write a short paragraph on the same topic in class, under normal conditions, and compare it to the submitted piece.

None of these steps require you to accuse the student of anything. They allow the student to demonstrate their ownership of the work - and if they cannot, that demonstration becomes a further strand of evidence.

Following School Policy and Escalating When Needed

Before taking any formal action, review your school's academic integrity policy. Schools vary widely in how they define and treat AI assistance, and some are still developing their approach. Some treat submitting AI-generated work as equivalent to plagiarism; others take an educational approach for a first incident, particularly at KS4.

If your concern is well-documented and the conversation with the student has not resolved it, involve your Head of Department before escalating to a formal process. Bring your documentation, the detection output, your comparison of prior work, and your notes from the student conversation. A shared professional judgment is stronger than an individual one, and it distributes the responsibility for what is ultimately a consequential decision.

Try GradeOrbit's AI Detection for Schools

GradeOrbit's AI detection feature is built to support professional judgment, not replace it. Every result includes a score against the pupil's own earlier writing, a confidence rating, a list of detected signals, and a plain-English reasoning paragraph - giving you something you can actually use in a documented case.

AI detection runs on school accounts, where pupils' work is held under the school's Data Processing Agreement and is never used to train AI models. Results appear on each pupil's profile, ready to sit alongside your own notes.

Start a free school trial and run your first AI detection check.