How Turnitin AI Detection Works: The Technical Explanation
Turnitin's AI Writing Indicator has flagged millions of student submissions since its 2023 launch. Most students know it exists. Fewer understand what it actually measures — which is why so many bypass attempts fail. This guide explains the precise signals Turnitin uses, why it's more reliable than most detectors, and exactly what needs to change in AI-generated text for it to pass.
Turnitin's Two Separate Systems
First, a critical distinction: Turnitin runs two completely independent systems on every submitted document. The Similarity Report (the traditional plagiarism checker) compares your text against Turnitin's database of academic sources and previous submissions. The AI Writing Indicator (launched April 2023) analyses the statistical properties of your prose to determine whether it was AI-generated.
These systems do not share signals. A document can score 0% on the Similarity Report and 90% on the AI Writing Indicator — meaning it contains no copied content but was clearly AI-generated. A document can score 40% similarity (lots of quoted sources without quotation marks) but 5% AI — meaning the writing is human-authored but poorly cited.
When people ask "does Turnitin detect AI?", they're usually asking about the AI Writing Indicator specifically. That's what this guide addresses.
The Four Primary Signals Turnitin Measures
Turnitin has not published full technical documentation for its AI Writing Indicator. The following is based on academic research into AI detection methodology, Turnitin's publicly available documentation, and empirical testing of the system's responses to specific text modifications:
1. Token-level perplexity
This is the foundational signal. Perplexity measures how surprised a language model would be by each word choice in context. AI language models generate text by choosing the statistically most probable next token — producing low-perplexity text. Human writers make surprising word choices, digress, use unusual metaphors, choose words that aren't the statistically obvious option.
WHY IT MATTERS
The more your text reads like the statistically expected output of a language model, the higher your AI score. This is why 'clear, well-structured writing' gets flagged — clarity at scale is a sign of optimisation, not intelligence.
Calculated at the sentence level, averaged across the document with weighting for outlier sentences.
2. Sentence-length burstiness
Burstiness measures the variance in sentence length across a passage. Mathematically, it's often measured as the coefficient of variation (standard deviation / mean) of sentence lengths across a document. Human writers produce high burstiness: some sentences are very short, some are long and elaborate, and the pattern of variation is irregular. AI models produce low burstiness: consistent medium-length sentences.
WHY IT MATTERS
A document where every sentence is 15-25 words is statistically suspicious. Real writing has one-word sentences. Real writing has 60-word constructions. The variation itself is human.
Measured at paragraph level and document level. Paragraph-level burstiness is weighted more heavily than document-level.
3. Transition pattern entropy
Turnitin tracks how varied the logical connections between sentences are across a document. Human writers use a wide range of transition strategies: implicit connection (no explicit link), contrast ('but', 'however'), addition ('and', 'also'), causation ('because', 'therefore'), temporal sequencing ('first', 'then'). AI text over-relies on a small set of transitions.
WHY IT MATTERS
A document using 'Furthermore', 'Additionally', and 'However' as the majority of transitions will score poorly on entropy. High transition diversity signals human authorship.
Calculated as Shannon entropy over the distribution of transition types. Low entropy (repetitive transitions) scores as AI-typical.
4. Structural regularity
This is Turnitin's more sophisticated signal that was added in later updates. It analyses whether paragraph structure follows a predictable pattern across the document — claim → evidence → elaboration, repeated with machine consistency. Human essays have structural variations: some paragraphs are primarily evidential, some are analytical, some are transitional. AI essays follow the same internal architecture throughout.
WHY IT MATTERS
If every paragraph in your document follows an identical internal logic, Turnitin's structural regularity signal will flag it regardless of how varied the vocabulary is.
Analysed at the paragraph sequence level. Documents with high structural regularity across 5+ consecutive paragraphs receive elevated AI scores.
What Turnitin's AI Percentage Actually Means
Turnitin reports the AI Writing Indicator as a percentage of the document that appears AI-generated. A score of 70% doesn't mean 70% of words came from AI — it means 70% of the text regions scored in the AI-typical range across the four signals above.
There's no universal threshold. Turnitin explicitly tells institutions not to use the score as a pass/fail determination — it's evidence to be considered alongside other factors. In practice, most institutions' academic integrity teams treat scores above 25-30% as requiring further investigation, and scores above 50% as strong evidence of AI use.
Turnitin also notes that its AI Writing Indicator was specifically calibrated to minimise false positives at the cost of some false negatives — meaning it's designed to be conservative. A document that scores 40% on Turnitin's indicator is more likely to be AI-generated than a document that scores 40% on GPTZero, because Turnitin's threshold for flagging is set higher. When Turnitin says AI, it typically means it.