How Turnitin AI Detection Works and Why It Flags Human Writing

A UC Santa Barbara student had an AI cheating accusation withdrawn after producing a hand-written notebook and browser history that showed the paper being written. The case is a useful lens for a specific question: how does Turnitin AI detection actually work, and why does it produce false positives on human writing like this student's?

What the UCSB case shows

The pattern is now familiar enough across campuses to describe without embellishment. A professor received a Turnitin AI similarity report flagging a student's essay. The student denied using AI, produced hand-written research notes with dated page numbers, and exported browser history showing library database access and source reading across the drafting period. The accusation was withdrawn. Related reporting on UCSB retraction cases describes the same arc at the same institution.

The question the case leaves open is the one students keep asking: if the student did not use AI, what was the detector responding to? The answer requires looking at what Turnitin's model actually measures.

How Turnitin's AI detector actually works

Turnitin's AI writing indicator is a classifier trained on examples of human-written and AI-generated academic text. It segments a submission into overlapping segments of a few hundred words and assigns each segment a probability that it was generated by a large language model. Those segment-level probabilities roll up into the percentage score that appears on an instructor's dashboard.

The signals the classifier learned during training are not disclosed in detail, but the research literature on AI text detection consistently identifies two statistical properties as the dominant features: perplexity and burstiness. Turnitin has described its model in general terms as pattern-based rather than rule-based, but the underlying mathematics is the same across detectors in this category.

Perplexity: how predictable the next word is

Perplexity measures how surprised a language model is by each word in a passage. If a model would have assigned high probability to the actual word that appears, perplexity is low. If the word is unexpected, perplexity is high. Large language models like GPT-4 are trained to produce high-probability sequences, so their output tends to have lower perplexity than typical human writing.

Careful human writers also produce low-perplexity text. A student revising a history essay to remove awkward phrasing, choose precise verbs, and conform to formal academic register is systematically lowering perplexity. The clearer the prose, the lower the perplexity, the higher the AI probability score. See our analysis of formal academic style and AI flags for how this pattern plays out in humanities writing.

Burstiness: how much sentence structure varies

Burstiness measures variation across sentences, in length, complexity, and perplexity itself. Human writing is thought to be more "bursty," mixing short sentences with long ones, simple structures with complex ones. AI output is often more uniform. Low burstiness combined with low perplexity is the signature detectors look for.

The problem is that uniformity is not unique to AI. Lab reports, legal memos, standardized-exam essays, and papers written to a strict rubric all tend to be structurally uniform. So does writing by students who were taught to construct parallel topic sentences or to keep sentence length under control.

Why human writing produces false positives

A detector trained on perplexity and burstiness will systematically misclassify categories of human writing that share statistical properties with model output. Peer-reviewed research has documented this across several populations.

The Stanford paper by Liang and colleagues (2023) tested seven GPT detectors on TOEFL essays written by non-native English speakers and found that more than half were classified as AI-generated. US-born eighth-grade essays run through the same detectors were almost always classified correctly as human. The gap is not about who used AI. It is about which human writers produce text that looks statistically like AI output.

Weber-Wulff et al. (2023), published in the International Journal of Educational Integrity, tested fourteen AI detection tools and concluded none performed reliably enough across conditions to be used as sole evidence in integrity proceedings. The populations most likely to be falsely flagged, based on the combined research, cluster around a few traits.

• Non-native English writers whose prose favors clearer syntax and higher-frequency vocabulary
• Students writing in formal academic registers (history, philosophy, legal writing)
• STEM students writing lab reports with constrained vocabulary and uniform structure
• Careful editors whose revision process reduces perplexity
• Writers using grammar tools like Grammarly that smooth sentence-to-sentence variation

What the UCSB student's evidence actually proved

The hand-written notebook and browser history did not rebut the detector's statistical analysis. They sidestepped it. A dated outline, source annotations in a notebook, and a browser history showing database searches during the drafting window are not probabilistic signals. They are a record of a process that produced the paper. That record is incompatible with the hypothesis that a language model generated the text.

This is the pattern worth understanding. The detector produced a number. The number came from signals that are also present in human writing. The way the case closed was not by challenging the math but by producing evidence from outside the document itself.

What this means if you have been flagged

A Turnitin AI score is a probability estimate from a classifier that has well-documented failure modes on human writing with low perplexity and low burstiness. It is not direct evidence of AI use. Most academic integrity policies require a preponderance of evidence or a higher standard. A single probabilistic flag, without corroboration, has not generally been treated as sufficient on its own in cases that go to review.

A written response should name the specific detector, cite the research on its false positive rates for writing like yours, and attach the process evidence you can preserve. If you want a letter that stitches those elements together without having to assemble them yourself, NotBot generates a personalized defense package built around the detector that flagged you and the writing process you can document. For broader procedural questions about what information you can request before a hearing, our procedural rights FAQ covers the ground.