Direct answer
Chatbot detectors work by scoring how statistically predictable a piece of text is, then classifying that score. They measure features such as perplexity (how surprising each word choice is) and burstiness (how much sentence length and structure vary), because large language models tend to produce text that is unusually smooth and low-variance. The output is a probability estimate with a confidence band, not a verdict — which is why every major detector, including Turnitin's, is documented as imperfect and unsuitable as standalone proof of misconduct [1].
What signals do chatbot detectors actually measure in a piece of writing?
Most detectors are built on the same core insight: a language model generates text by repeatedly selecting the most probable next token, so text that came from a language model tends to sit in the high-probability region of that distribution. Detectors therefore compute a likelihood score over the passage and threshold it into a label. OpenAI's own AI text classifier was built exactly this way, and when OpenAI retired it, the company disclosed that it correctly identified only 26% of AI-written text as "possibly AI-written" while incorrectly flagging 9% of human-written text [2]. That asymmetry — a weak true-positive rate paired with a non-trivial false-positive rate — is the single most important fact about how these systems behave.
Beyond raw token likelihood, detectors aggregate secondary features that correlate with machine generation. Sentence-length uniformity, low lexical diversity, repeated syntactic templates, and an absence of the small irregularities that human writers introduce (contractions, asides, uneven paragraph rhythm) all push a passage toward the "AI" side of the classifier. Turnitin's detector is documented as analyzing patterns of perplexity and burstiness at the segment level and then rolling those segment signals into a document-level estimate [1]. The score you see is therefore a composite, not a measurement of any single property.
It is also worth being precise about what the number means. A detector does not locate a source, verify authorship, or prove that a chatbot produced the text; it reports how closely the text's statistical profile resembles the profile of model-generated writing it has seen before. OpenAI noted that detection quality degrades sharply on text outside the distribution the classifier was trained on, which is precisely why results shift between disciplines, genres, and writing styles [2]. Treat the score as a similarity measure between your prose and a model's prose, and the rest of the detector's behavior becomes predictable.
Why do chatbot detectors give false positives on human writing?
False positives are not a bug at the margins; they are a structural consequence of the signal detectors rely on. Any human writing that is formulaic, heavily templated, or written under a constraint that suppresses stylistic variation will look statistically similar to machine output. Reporting on detector bias found that AI detectors flagged a large share of TOEFL essays written by non-native English speakers as AI-generated while flagging almost none of the essays written by US-born students [3]. The essays were human-written. The detectors were not detecting AI; they were detecting a low-variance writing style that happens to overlap with the model's signature.
The same mechanism explains false positives in other populations. Students writing to a rigid rubric, researchers following a strict methods-section convention, and anyone drafting in a second language under time pressure all produce prose with reduced perplexity and compressed burstiness. Because the classifier has no access to your drafting history, your notes, or your revision trail, it cannot distinguish "written by a person who writes in a regular way" from "written by a model that writes in a regular way" [3]. The only evidence it has is the text itself.
This is why detector guidance consistently frames a high score as a prompt to investigate rather than a conclusion. Turnitin states plainly that its AI writing detector is not foolproof and advises institutions against using the indicator as the sole basis for an adverse decision [1]. When a detector's confidence is low, the platform surfaces a *% marker instead of a precise percentage, acknowledging that the underlying estimate is too weak to report as a number [1]. A score is an opening question, and the answer to that question is a conversation about process, drafts, and sources — not an automatic penalty.
How can I see the same AI and similarity report my instructor sees before I submit?
The most practical defense against a surprise flag is to look at the report yourself first. The AI writing report presents a document-level percentage alongside a segment-level breakdown and highlights the specific passages the model considers AI-generated, which is the same view an instructor opens in the learning management system [4]. Seeing that view changes the conversation: instead of arguing about an abstract number, you can point to the exact sentences that triggered the signal and explain how they were written.
There is one structural catch. A report is generated against a specific submitted file, so there is no way to view a report for a draft you have not run through a check [4]. That means pre-submission checking is the only route to seeing your own detection output before the version that counts is submitted. Turnitin's own documentation frames the report as a starting point for discussion rather than a final judgment, which is exactly how a preview is best used [4].
Understanding the mechanism also tells you what a preview is good for. Because detectors respond to low perplexity and compressed burstiness, a preview reveals which passages are statistically exposed — the flat, uniform stretches that a classifier reads as machine-like. You can then decide whether those passages genuinely need revision, or whether they are simply the product of a formal register that your discipline expects. Either way, you are making the decision with the same evidence your instructor will see, rather than discovering it after the deadline [4].
Knowing the mechanism is only half of it — the other half is seeing how the classifier actually reads your own draft. turnitin0 runs a pre-submission check that returns the real Turnitin AI detection report and similarity report together, so you can inspect the document-level score, the segment-level breakdown, and the highlighted passages before anything is final. It is the same report format your instructor opens, delivered in minutes, without adding your file to the student paper database.
※ Turnitin0.com - Actual Turnitin AI Report Cover, Score, Flag And Similarity Summary
FAQ
Do chatbot detectors actually prove a text was written by AI?
No. Detectors output a probability estimate derived from statistical features such as perplexity and burstiness, and Turnitin explicitly documents its AI writing detector as not foolproof and unsuitable as the sole basis for an adverse decision [1]. A score indicates resemblance to model-generated writing, nothing more.
Why does the same essay get different AI scores on different detectors?
Each detector uses a different model, training corpus, and threshold, so the same text can land on opposite sides of a cutoff. OpenAI's classifier, for example, identified only 26% of AI-written text while flagging 9% of human-written text, and its accuracy dropped on text unlike its training data [2]. Variation between tools is expected, not evidence that one is lying.
Can a high AI score happen to someone who wrote everything themselves?
Yes, and it happens most often to writers whose prose is formulaic or constrained. Detectors flagged a large share of human-written TOEFL essays by non-native English speakers as AI-generated while flagging almost none of US-born students' essays [3]. Low stylistic variance, not AI authorship, is what triggered the signal.
What exactly does the Turnitin AI writing report show?
It shows a document-level AI percentage, a segment-level breakdown, and highlighted passages the model considers AI-generated — the same interface instructors see in their LMS [4]. When confidence is low, Turnitin displays *% instead of an exact figure [1].
Can I see my own AI and similarity report before I submit?
Yes, but only by running a check on the draft, because a report is generated against a specific submitted file [4]. A pre-submission check returns the AI detection report and similarity report together so you can review the flagged segments and decide what, if anything, to revise.