Turnitin detects AI writing with a machine-learning classifier. It analyzes eligible prose in overlapping sections, assigns probability-like scores, pools those results at sentence level, and calculates how much qualifying text appears likely to be AI-generated or AI-altered.
The short answer
That means the system does not need to find the exact prompt, locate the answer in ChatGPT, or match the text to a stored AI response. Instead, a model trained to distinguish examples of human and AI writing evaluates patterns across pieces of the document.
Turnitin does not publicly disclose every model feature or decision threshold. Claims that it relies on one tell—such as a favorite word, sentence length, “perplexity,” or an invisible watermark—oversimplify what the company describes as a segmented, aggregated prediction process.
The five stages of Turnitin AI detection
It checks whether the file qualifies
The current AI Writing Report generally requires a supported file type, less than 100 MB, 300–30,000 words of long-form prose, and a supported language. A file can receive no result if it cannot be processed.
It extracts qualifying prose
The model focuses on prose sentences in paragraphs—the type of writing found in essays, articles, and dissertations. Code, poetry, scripts, tables, bullet lists, and annotated bibliographies are not reliably handled in the same way.
It creates overlapping text segments
According to Turnitin, extracted sentences are grouped into overlapping sections for prediction. The overlap gives the classifier surrounding context instead of forcing it to judge each sentence in isolation.
It scores and pools the segments
Each segment receives a value from 0 to 1 representing how human-like or AI-like the model finds it. Sentences inherit their segment’s score; where segments overlap, a sentence may receive several scores that are pooled into one.
It aggregates the document result
Sentence results are combined into the overall AI-writing percentage. Current reports highlight qualifying text considered likely AI-generated or AI-generated and subsequently altered by an AI paraphraser or bypasser.
A sentence near a boundary can be evaluated in more than one context. Those multiple scores are pooled before the document-level result is calculated.
What writing signals does the model notice?
Turnitin publicly explains the scoring pipeline but does not publish a complete checklist of its internal features. At a high level, AI classifiers learn statistical differences between human and machine-generated language across context: how words, structures, and sequences tend to occur together.
It is reasonable to discuss general concepts such as predictability, repetition, syntactic regularity, and variation when explaining AI detectors broadly. But none of these is a magic trigger, and Turnitin does not say that one phrase or score automatically causes a flag.
Why context beats a “banned words” list
Words like “delve,” “moreover,” or “tapestry” are sometimes called AI giveaways online. Humans use them too. A classifier evaluates broader patterns across a segment; deleting a few fashionable words does not reliably change the underlying pattern or establish authentic authorship.
Can it detect paraphrased AI?
Turnitin says its English detector can identify text likely generated by an LLM and then modified by an AI paraphraser, bypasser, or word spinner. Support varies by language, and detection remains probabilistic. A highlight means “review this text,” not “this transformation has been proven.”
Four common myths about Turnitin AI detection
The AI-writing feature classifies text patterns. Database matching belongs to the separate Similarity Report.
The score represents the share of qualifying prose identified as likely AI-written or AI-altered—not confidence in guilt.
Segment-based analysis looks beyond isolated vocabulary, and English reports can include likely AI-paraphrased text.
Turnitin explicitly says its model can misidentify text and should not be the sole basis for adverse action.
For a detailed explanation of visible percentages, see our companion guide: Turnitin AI detection scores from 0 to 100.
Why the detector can be wrong
Classification is an estimate. Human writing can resemble patterns in the model’s AI examples, and revised AI writing can resemble human examples. Topic, genre, language, length, editing, and formatting all affect the available signal.
| Limitation | Why it matters |
|---|---|
| False positives | Fully human prose can be misclassified. Current reports therefore suppress exact 1–19% results and display *%. |
| False negatives | AI-generated writing may not be identified, particularly after substantial, meaningful human revision. |
| Non-prose content | Lists, code, poetry, tables, and other unconventional formats may not count as qualifying text. |
| Changing systems | Generative models and detectors evolve. A result reflects a particular model version and submission time. |
Turnitin reports that for documents with more than 20% likely AI-generated content, its document-level false-positive rate is below 1%. That is a vendor-reported performance figure, not a promise that any individual highlight is correct. Turnitin still instructs educators to combine the report with human judgment and institutional policy.
How students and educators should review a result
- Open the full AI Writing Report. The overall percentage alone removes the passage-level context needed for a fair review.
- Confirm what counted as qualifying text. The percentage may cover less than the whole document.
- Compare the highlights with the writer’s process. Check notes, outlines, sources, document history, earlier drafts, and approved AI disclosures.
- Discuss the ideas. Ask the writer to explain an argument, evidence choice, or revision. Understanding cannot be reduced to a classifier output.
- Apply the actual course policy. Detection is not the same as prohibited use, especially where brainstorming, grammar help, or disclosed assistance is permitted.
Students should preserve revision history and write from their own notes and sources. If a human-written passage is questioned, process evidence is more useful than repeatedly rewriting authentic prose to satisfy different detectors.
Check your draft with PaperCheck
PaperCheck is an accurate, free AI-writing detector for reviewing a draft before submission. It helps identify passages worth rereading while keeping the decision in your hands. Different detectors use different models, so it cannot duplicate or guarantee a Turnitin score.
PaperCheck is designed around data minimization: submitted text is processed in real time and deleted immediately after analysis, while results may be cached only for the current session, as described in our privacy policy.
Frequently asked questions
Does Turnitin detect ChatGPT, Gemini, and Claude?
Turnitin describes its system as detecting text likely produced by large language models rather than identifying a guaranteed source model for every passage. Coverage evolves as models change, so a highlight should not be treated as proof that one named tool was used.
Does Turnitin use perplexity?
Predictability-related concepts are common in AI detection research, but Turnitin does not publish a complete list of proprietary model features. Its official explanation focuses on overlapping segments, 0–1 classifications, pooled sentence scores, and document-level aggregation.
Can Turnitin detect AI after paraphrasing?
Its English detector is designed to identify likely AI-generated text that may have been altered with AI paraphrasers, bypassers, or word spinners. No detector can guarantee detection in every case.
Does a 0% score prove that a person wrote everything?
No. It means the model did not identify qualifying prose as likely AI-generated or AI-altered in that scan. It is not an authorship certificate.
Is PaperCheck affiliated with Turnitin?
No. PaperCheck is an independent alternative for free, privacy-conscious pre-submission review. Results will not necessarily match Turnitin because the tools use different models.