Direct answer
No single tell proves a passage came from ChatGPT, and no detector — including Turnitin's — issues a verdict; they issue a probability over specific segments. The reliable approach is to combine textual evidence (uniform rhythm, formulaic transitions, thin sourcing) with a segment-level AI writing report, then treat the score as a signal that needs interpretation rather than a confession [1]. Turnitin only scores prose segments of roughly 300 words or more and reports the share of qualifying text its model flags, which is why short passages can look either innocent or damning for the wrong reasons [3].
What Signals Actually Reveal That a Piece of Text Was Written by ChatGPT?
Surface tells are weak individually but meaningful in clusters. ChatGPT-style output tends toward low sentence-length variance, heavily signposted transitions ("Moreover," "In conclusion"), balanced "on the one hand / on the other hand" framing, and confident summary sentences that restate rather than advance the argument [2]. None of these are proof: competent academic writers are trained to produce exactly that register, and a human editing an AI draft usually strips the most obvious markers while leaving the underlying uniformity intact [2].
The stronger evidence is forensic rather than stylistic. Look for claims that cannot be traced to a source, citations that exist but do not say what the text claims, dates or statistics that are subtly wrong, and paragraphs that never engage a specific counter-argument. AI-generated prose is fluent at the level of the sentence and thin at the level of the evidence chain, so a paragraph that sounds polished but collapses under one "where does this come from?" question is more diagnostic than any single phrase [2].
Context also changes the reading. A 150-word discussion post and a 4,000-word dissertation chapter carry different baselines, and the same stylistic features that look machine-like in a personal reflection look normal in a literature review. That is why reviewers are advised to judge text against the genre's conventions and the writer's own prior work rather than against a generic list of "AI words" [2].
How Does Turnitin's AI Detector Decide Whether Text Is AI-Generated, and What Does Its Score Mean?
Turnitin's AI writing detection works at the segment level rather than the document level. It evaluates prose in windows of roughly 300 words, and the headline percentage is the proportion of qualifying text its model classifies as AI-generated — so a report showing 34% means about a third of the scored prose was flagged, not that a third of the essay is fake [3]. Reference lists, quotations, and very short submissions fall outside the scored range, which is a common source of confusion when a heavily quoted paper returns a low or blank score [3].
The percentage is also displayed unevenly by design. When the model's confidence sits below its 20% threshold, Turnitin shows *% instead of an exact figure, signalling a low-confidence result rather than a clean pass [3]. That distinction matters enormously in practice: *% is not the same as 0%, and treating it as proof of innocence — or as proof of guilt — misreads what the report actually says [3].
Most importantly, Turnitin frames the report as an indicator that supports human judgement, not as an accusation. False positives occur with non-native English writers, heavily templated genres, and text that has been through multiple rounds of grammar editing, and the company's own guidance tells instructors to combine the score with context and conversation [3]. If you are the writer, that is the frame to hold onto: a flagged score opens a discussion about how the text was produced, and documentation of your drafting process is the strongest thing you can bring to it [3].
Can You Check a Draft for AI Writing Before You Submit It?
Yes — pre-submission checking exists, and the main thing to understand is that not all checks are equivalent. A non-repository check compares your file against the databases without adding it to the student paper repository, so your draft is not stored as a future match against other students, and the resulting reports are yours to review and delete [4]. That property is what makes previewing a draft materially different from submitting it to your institution's assignment inbox [4].
The practical value is timing. Seeing a segment-level AI report while you can still revise lets you act on the specific paragraphs that were flagged — rewriting them in your own voice, adding the source engagement that was missing, or replacing generic framing with your own analysis — instead of discovering the problem after a grade has been recorded [4]. It also calibrates your expectations: if your draft returns *%, you learn that the signal is low-confidence before an instructor ever sees it [3][4].
One caveat belongs here rather than in a footnote. Institutional policies differ on whether students may run their own checks, and some departments treat pre-submission checking as acceptable preparation while others regard repeated checking as an integrity issue in itself. Read your programme's academic-integrity guidance first, and use any preview tool as a revision aid rather than a gate to be gamed [4].
If you want that preview to look exactly like the report your marker will open — segment percentages, flagged passages, and the similarity summary side by side — turnitin0 produces the real Turnitin AI detection and similarity reports on your draft, non-repository, typically in under 15 minutes, with 100,000+ reports delivered to 20,000+ students.
※ Turnitin0.com - Actual Turnitin AI Report Cover, Score, Flag And Similarity Summary
FAQ
Can you tell if text is from ChatGPT just by reading it?
Not reliably. Clusters of formulaic transitions and uniform sentence rhythm raise suspicion, but trained writers produce the same features and edited AI text hides the obvious markers [2].
What does a *% score on a Turnitin AI report mean?
It means the model's confidence fell below Turnitin's 20% threshold, so no exact percentage is shown. It is a low-confidence signal, not a clean zero [3].
Why did my AI score change between two checks of the same essay?
Scores are computed over ~300-word scored segments, so edits that shift paragraph boundaries, add quotations, or change length can move the percentage even when the substance is unchanged [3].
Is it safe to check my draft before submitting it?
With a non-repository check, the file is not added to the student paper database and you can delete it afterwards. Check your institution's policy on pre-submission checking first [4].
Does a high AI score mean I will be penalised?
Not automatically. Turnitin positions the report as an indicator for human review, and institutions are expected to weigh it alongside drafting evidence and context [1][3].