GradeOrbit vs Turnitin: How AI Detection Scoring Actually Works

If you teach in a UK secondary school, the question of how to handle AI-generated student work has probably landed on your desk in the past year. Two names tend to come up: Turnitin, which has bolted AI detection onto its long-standing plagiarism platform, and GradeOrbit, which built AI detection into a marking-first product from the start.

This post explains how the underlying scoring works in plain English, why the two tools give different numbers on the same piece of work, and how to use a likelihood score as one piece of evidence rather than a verdict. The aim is not to crown a winner - it is to help you make an informed call when ChatGPT or Claude shows up uninvited in a Year 11 essay.

What an AI Detection Score Actually Is

Every reputable AI detection tool - including GradeOrbit and Turnitin - produces a probabilistic output. With most tools, the number you see is a model's estimate of how strongly the text resembles patterns associated with AI generation; GradeOrbit's number instead estimates how far the work departs from the pupil's own earlier writing. Either way, it is not a forensic test result. It is closer to a weather forecast: useful, directional, sometimes wrong.

That distinction matters because the way the score is presented shapes how teachers and senior leaders react to it. A bare percentage with no context invites a binary interpretation. A score paired with reasoning, detected signals, and confidence labelling invites professional judgement. The difference between those two experiences is what separates a useful classroom tool from a disciplinary risk.

How Turnitin Approaches AI Detection

Turnitin's AI detection sits inside its broader similarity report. It compares submitted writing against statistical patterns learned from large samples of human and AI-generated text, then returns a percentage describing how much of the submission appears AI-generated. Turnitin has been transparent that its model can produce false positives, particularly on writing from English-as-an-additional-language students and on short passages.

The product makes sense for institutions already using Turnitin for plagiarism. It is also tied to a wider workflow - submission portals, originality scores, similarity matches against archived student work. That archival side is part of why some schools are cautious about it: student work is processed and, in some cases, retained by the platform.

How GradeOrbit's AI Detection Works

GradeOrbit takes a different path. The detection tool sits inside a marking-first product, so it is built for the same moment teachers reach for it in real life - a piece of work has just arrived and you want a second opinion before you mark it.

Detection is part of GradeOrbit school accounts. Rather than producing a generic AI-likelihood score, it compares a pupil's new work against that same pupil's own earlier work already on file. Teachers run it from a pupil's profile or on a whole class at once, uploading scans, photos or PDFs (or using the phone camera link); for a class set, bulk upload matches the name on each cover page to the class list. The tool then returns four things together:

• A score from 0 to 100 measuring how far the new work departs from the pupil's own earlier writing (low = consistent with their writing, high = differs in ways worth a closer look).
• A confidence label: Low, Medium, or High - describing how sure the model is.
• A list of detected signals: the specific patterns it noticed.
• A short reasoning paragraph in plain English.

Detection is included in every school teacher seat, with no per-check charge and no model choice to configure. A pupil with no earlier work on file is skipped, and the teacher is told, rather than being given a generic score.

Crucially, AI detection runs on school accounts, where pupils' work is held under the school's Data Processing Agreement - that earlier writing on file is what the comparison uses. Work is wiped at the end of the school's chosen retention window or at offboarding, which gives schools a clear, written answer to data-retention questions.

Why the Same Essay Can Score Differently

If you run the same paragraph through Turnitin and GradeOrbit, the numbers will almost never match exactly. That is not a bug - it is a property of how these systems work.

Different detectors are trained on different samples of human and AI writing. They weight different signals - sentence rhythm, vocabulary choice, paragraph structure, predictability of the next word - in different combinations. A 64% score in one tool and a 28% in another does not mean one is right and one is wrong. It means the two models found different amounts of evidence in the same text.

This is also why no single number should ever be treated as proof of misconduct. Teachers' colleges, exam boards, and the JCQ have all been clear: AI detection results are corroborating evidence, never a verdict on their own.

Using a Likelihood Score Well

The score is most useful when you use it to structure your professional judgement rather than replace it. A practical workflow looks like this:

Anchor on the Student's Baseline

Your own knowledge of the pupil is the strongest signal you have. If the writing feels stylistically alien to anything they have produced in class, that is meaningful regardless of what any tool reports. A high likelihood score that aligns with your gut tells you to dig in. A high score that does not match your knowledge of the student tells you to be careful and look closer before reacting.

Read the Reasoning, Not Just the Number

GradeOrbit's reasoning paragraph and detected-signals list exist precisely so a 78% score is not a black box. If the signals say "low burstiness, uniform sentence length, generic transitions," you can hold that against the actual essay in front of you. Sometimes the signals are convincing. Sometimes they describe a Year 10 who simply writes in clean, even sentences.

Have the Conversation Before the Sanction

If detection scores and your professional read both point the same way, the next step is almost always a short conversation with the student. Ask them to talk through their argument, defend a choice of evidence, or write a short paragraph on the same topic on paper. The principle holds across every detection product on the market: the score opens the door, the conversation closes the case.

Where Each Tool Fits

Turnitin makes sense for schools already deep in its ecosystem, particularly where coursework is centrally submitted and plagiarism similarity is a core part of the workflow. GradeOrbit makes sense for schools that want AI detection as part of the same place they mark - and who want pupil work held under the school's own Data Processing Agreement with a set retention window.

Other tools exist too. None of them, used alone, will give you certainty. All of them are useful when treated as a structured second opinion rather than a judge.

The Honest Limit of All AI Detection

Detection technology is in a race with generative models that are getting better at sounding human. A student who runs AI output through a paraphraser, edits it by hand, and adds a personal anecdote will drop their likelihood score on any tool currently available. That is the reality every detection product - including GradeOrbit's - has to be honest about.

What this means in practice is that detection tools work best as part of a wider academic-integrity culture: clear policies, in-class writing samples, regular drafting checkpoints, and conversations with students about what AI use is and is not acceptable. The tool gives you signal. The school's culture turns signal into integrity.

Try GradeOrbit's AI Detection

If you want to see how the score, confidence label, detected signals, and reasoning paragraph work together on real student writing, GradeOrbit's AI detection is included in every school teacher seat, and pupils' work is held under the school's Data Processing Agreement. See how detection works on a school account and start a free school trial before deciding whether it fits your marking routine.