Since Turnitin rolled out its AI writing detection feature in 2023, it has become standard infrastructure at the majority of UK universities. If you submit an essay through Turnitin — which most UK undergraduates do — your work is being analysed for AI patterns every single time, whether your lecturer looks at the score or not.
Understanding how this system actually works gives you a significant advantage. Most of the anxiety students feel about AI detection comes from not knowing what the tool is looking for.
How Turnitin's AI detection actually works
Turnitin's AI detection uses a statistical language model to assess the probability that each sentence was generated by AI. It doesn't compare your essay to a database of AI-generated text the way plagiarism detection compares to existing sources. Instead, it analyses the writing patterns themselves.
The model evaluates text based on two core signals:
- Perplexity — how predictable each word choice is given the surrounding text. AI language models generate text by selecting statistically likely next tokens. This makes AI writing smooth and coherent but also statistically predictable. Human writers make more surprising word choices.
- Burstiness — how much sentence length varies throughout the text. Human writing has natural rhythm variation — short punchy sentences followed by longer elaborated ones. AI writing tends to produce sentences of similar length throughout, creating a uniform rhythm that detection models identify.
Turnitin's model was trained on a large corpus of both human-written and AI-generated academic text. It outputs a probability score — the percentage of text it believes was AI-generated — which appears alongside the standard similarity report in the lecturer's dashboard.
Turnitin's AI score and its plagiarism similarity score are completely separate. A high AI score doesn't mean your work was copied from another source — it means the writing patterns statistically resemble AI-generated text. A student can have a 0% similarity score and a 90% AI score simultaneously.
What the AI score means — and what triggers a review
Turnitin reports the percentage of text it believes is AI-generated. There is no universal threshold at which action is taken — each institution sets its own policy. However, based on publicly available guidance from UK universities, the general picture is:
These thresholds are indicative. Your university may act at a lower threshold or have a different process entirely. The module handbook or your institution's academic integrity policy is the definitive source.
How accurate is it really?
Turnitin has claimed accuracy rates above 98% in its own testing. Independent research has produced more mixed results, and the figures differ significantly depending on what type of text is being tested.
Where Turnitin performs well:
- Raw, unedited output from major AI models (ChatGPT, Gemini, Claude)
- Essays where the entire text is AI-generated with no human editing
- Heavily formulaic AI output that closely follows the model's default style
Where Turnitin struggles:
- AI text that has been substantially rewritten or edited by a human
- Human text that happens to be written in a uniform, formal style
- Non-native English writers whose writing patterns overlap with AI output
- Subject-specific writing conventions that produce low burstiness naturally (law, science, nursing)
Who gets false positives
False positives — human-written essays flagged as AI-generated — are a real and documented problem with Turnitin's detection. The groups most at risk are:
- Non-native English speakers — students writing in their second or third language tend to use simpler, more uniform vocabulary and sentence structures. These patterns overlap with AI output and can produce elevated scores on entirely human-written work.
- Students in highly formulaic subjects — law problem questions (IRAC structure), nursing reflective essays (Gibbs cycle), scientific lab reports, and case study analyses all follow prescribed structures that reduce natural variation.
- Students who write formally — academic writing conventions include hedging language, formal transitions, and standardised argument structures. Heavy use of these can look AI-like even when the writing is entirely original.
- Students who edit heavily — extensive editing and polishing can smooth out the natural variation in sentence structure, raising the statistical uniformity of the text.
UK universities are aware of the false positive problem. Most have guidance stating that a high AI score alone is not sufficient evidence of misconduct — it triggers a review process, not automatic punishment. But being prepared matters.
What Turnitin AI detection misses
Understanding what Turnitin doesn't catch is as important as understanding what it does. There are categories of AI use that the current version of Turnitin's detection largely misses:
- AI used for research only — using ChatGPT to understand a concept, explore counterarguments, or generate a reading list, then writing the essay yourself. The final text is human-generated and will read that way.
- AI-assisted editing — using AI to improve grammar, vary sentence structure, or suggest alternative phrasings. If you then review and apply changes selectively, the resulting text typically has sufficient human variation to score low.
- Heavily rewritten AI text — if substantial portions of AI output have been rewritten in the student's own voice, the detection signal degrades significantly.
- AI for brainstorming and outlining — generating an essay structure, argument list, or counterargument summary, then writing all the actual prose from scratch.
Just because Turnitin can't detect certain types of AI use doesn't mean those uses are permitted. Your module's policy may prohibit AI assistance even where it's undetectable. Always check your specific policy — detection capability and permitted use are separate questions.
What to do before you submit
Whether or not you've used AI, running a pre-submission check is simply good practice. Turnitin's score appears in your lecturer's report — you don't see it — so the first time you know there's a problem is after you've already submitted.
What SafeGrade checks that Turnitin looks for:
- Perplexity and burstiness across your full essay
- AI-typical phrase patterns (specific phrases that appear disproportionately in generated text)
- Sentence length variation throughout the document
- Vocabulary diversity across the whole essay
- Paragraph structure uniformity
- Fabricated reference patterns (AI-generated citations that don't exist)
Running SafeGrade's Writing Analysis is free and unlimited — it gives you the six dimensions Turnitin cares about before you submit. If anything looks concerning, the AI Risk Check (Deep Scan) goes further and returns a specific AI risk verdict. One free per month.
It takes 30 seconds.