Direct answer
AI-generated text is identified through clusters of stylistic signals — em-dash overuse, "it's not just X, it's Y" constructions, rule-of-three lists, canned transitions such as "Moreover" and "Overall," inflated claims about significance, and uniform sentence rhythm — rather than any single tell. Wikipedia editors maintain a public catalogue of these markers in the Signs of AI writing guidance page, but that page is explicit that the markers are probabilistic evidence, not proof [1]. Any serious diagnosis combines stylistic reading with structural checks on citations and edit history, and detection tools add a statistical layer on top [2].
What are the reliable signs that a piece of text was written by AI?
The most reliable signs are combinations of habits, not isolated quirks. Wikipedia's editor-maintained catalogue groups them into formatting tells (heavy boldface, decorative bullet lists, headings that restate the paragraph), phrasing tells (em-dash chains, "not just X but Y," rule-of-three triads, reflexive "Overall" and "Moreover" openers), and content tells (vague claims that a topic "plays a crucial role" or "reflects broader trends" without a citation behind them) [1]. A single em-dash proves nothing; four of these patterns inside two paragraphs is a different matter.
Vocabulary is the weakest signal on its own. Words like "delve," "tapestry," "testament," and "landscape" are frequently cited as AI markers, but human writers in academic and journalistic registers use them too, and newer models have been tuned away from the most notorious ones [1]. Treating a word list as a detector produces exactly the false accusations that Wikipedia editors warn against, which is why their guidance frames every marker as a prompt to look closer rather than a verdict [1].
Structural and provenance checks are stronger than style. Wikipedia editors routinely verify whether the cited sources exist, whether they say what the text claims, and whether the article's edit history shows a single large paste from a new account [1]. Turnitin's own explanation of detection follows the same logic from the other direction: its model does not look for keywords but measures how statistically predictable each token is, then segments the document and scores sentence-level likelihood, because machine text tends to be more predictable and more rhythmically uniform than human prose [2]. That is why a paraphrase that removes the obvious vocabulary tells can still be flagged — the underlying regularity survives [2].
Why do Wikipedia editors treat AI-writing detection as unreliable, and what do their guidelines actually say?
Wikipedia's guidance is deliberately cautious because the cost of a wrong accusation is high. The Signs of AI writing page states that its indicators are not definitive and that editors should not label contributions as AI-generated on style alone; the page functions as an investigative aid for discussions, not a scoring rubric [1]. Editors are directed toward behavioural and sourcing evidence — fabricated citations, sources that do not support the sentence, abrupt wholesale rewrites — because those are verifiable in a way that "this sounds like ChatGPT" is not [1].
Commercial detectors inherit the same limitation, and Turnitin says so plainly. The company's own documentation notes that its AI writing report may not always be accurate, must be interpreted with human judgement, and is not definitive proof of misconduct [3]. It also explains the asterisk convention: when the detected share falls below the report's confidence threshold, the interface displays an asterisk rather than an exact percentage, signalling a low-confidence result rather than a clean bill of health or a firm accusation [3].
The practical takeaway is that detection is a spectrum with a wide grey zone. Text that has been lightly edited by a human, text drafted in a second language, and text that is highly formulaic by genre — lab reports, literature reviews, policy summaries — all sit in that zone, which is why both Wikipedia editors and Turnitin push the decision back to a human reader with context [1][3]. A detector score is evidence to be explained, not a verdict to be enforced [3].
How can a student verify whether their own draft will be flagged as AI-generated before submitting?
The verification step is to reproduce the conditions under which the draft will actually be judged. Turnitin's score reflects the share of qualifying prose in the submitted document that its model classifies as AI-generated, and short documents or text dominated by quotations and references may be excluded from the calculation entirely [4]. That means a student who tests an excerpt, or tests a version with different formatting, can get a materially different result from the one their instructor will see — the full document, in the submitted file format, is what matters [4].
Access is the second constraint. Students generally see an AI writing report only when their instructor has enabled it for the assignment, so the report is not reliably available as a self-check inside the institutional system [4]. The workaround that actually answers the question is a pre-submission check that returns the same report an instructor would see, on the same document, before the deadline rather than after it.
Reading the report correctly matters as much as obtaining it. A flagged percentage is a proportion of qualifying text, not a probability that the student cheated, and the sentence-level highlighting is where the useful information sits — it shows which passages triggered the model and whether those passages are the ones the student actually wrote with assistance [4]. Students who review the highlighted spans, revise the genuinely machine-drafted sections, and re-check get a far more actionable answer than students who fixate on the headline number [3][4].
If you have read this far, you already know the uncomfortable part: stylistic markers tell you that something might be machine-drafted, but only a report on your actual document tells you what an instructor's system will conclude. turnitin0 exists for that gap — it returns the Turnitin AI detection report and the similarity report together, on your own draft, in the format your institution uses, so the grey zone becomes a specific set of highlighted sentences you can revise before the deadline instead of a guess you defend afterwards.
※ Turnitin0.com - Actual Turnitin AI Report Cover, Score, Flag And Similarity Summary
FAQ
Is there a definitive way to prove text was AI-generated?
No. Wikipedia's own guidance describes its indicators as probabilistic and warns against labelling text as AI-generated on style alone [1], and Turnitin states its report is not definitive proof of misconduct [3]. The honest answer is that you can build a strong circumstantial case, not certainty.
Do em-dashes and the word "delve" mean a text is AI?
Not on their own. Those are popular shorthand markers, but they appear in plenty of human academic and journalistic writing, and newer models use them less [1]. They are worth noting as one signal among several, never as a standalone verdict.
Why does Turnitin show an asterisk instead of a percentage?
The asterisk appears when the detected share of AI-generated text falls below the report's confidence threshold, which signals a low-confidence result rather than a confirmed finding [3]. Treat it as "the model is not confident here," not as either a pass or a fail.
Can I check my own draft before I submit it?
Institutional reports are usually visible to students only when an instructor enables them for the assignment [4]. A pre-submission check on the full document, in the file format you will submit, is the practical way to see the same report beforehand.
Does editing AI text by hand remove the signals?
Partly. Rewriting vocabulary and sentence rhythm can lower a score, but the deeper statistical regularity that detectors measure sometimes survives light editing, which is why a re-check after revision is the reliable test [2][4].