Direct answer
Direct Answer - ** No single sentence, paragraph, or percentage can prove that a piece of text was written by AI. The honest answer is a probability, not a verdict: detectors estimate how closely your wording matches patterns typical of machine-generated prose, and that estimate has to be read alongside the document's context [1]. Treat any "AI score" as a signal that tells you where to look, never as evidence of misconduct on its own [1]. What you can do is check the specific passages a detector flags, understand why they were flagged, and confirm your draft's status before it reaches a marker.
How Can You Tell Whether a Piece of Writing Was Generated by AI?
The most reliable approach is layered: read for stylistic tells, then verify with a detector, then weigh both against what you know about how the text was produced. Turnitin's own guidance frames its AI indicator as an estimate of the likelihood that text was AI-generated, explicitly not proof that a student cheated, and it asks staff to use the result as a starting point for a conversation rather than a conclusion [1]. That framing matters because the tool is doing statistical pattern-matching, and pattern-matching is exactly the kind of judgement that fails on unusual but entirely human writing.
Stylistic tells are real but weak on their own. Human readers notice uniform sentence length, generic transition phrases, confident generalities with no named sources, and paragraphs that summarise a topic without ever taking a position. Those features show up in AI output often enough to be useful, but they also show up in tired student writing at 2 a.m., in translated prose, and in any essay written to a rigid template. Reading for tells is a way to decide where to look, not a way to reach a conclusion.
The harder problem is that most suspicious text is now mixed. A student may outline with a model, write the body themselves, and ask for a polish pass on two paragraphs — a workflow that produces a document no single label describes well. Because of that, the strongest evidence of authorship is usually process evidence rather than textual evidence: dated drafts, notes, version history, and the ability to explain your own reasoning in a viva or supervision meeting. If you are the author, keep those artefacts; if you are the reader, ask for them before you accuse anyone [1].
Finally, be precise about what you are actually asking. "Is this AI?" can mean "did a model generate this text", "did a model write this sentence", or "was this assignment produced without the student's own work". Those are three different questions with three different standards of proof, and conflating them is the most common reason people reach the wrong conclusion [1].
Why Do AI Detectors Flag Human Writing as AI, and How Reliable Are the Percentages?
Detectors flag human writing because they are not reading for meaning — they are scoring how predictable each token is given the ones before it. Formulaic academic register, heavy use of stock connectives, and grammatically impeccable but semantically thin sentences all sit close to the statistical centre of machine writing. That is why careful writers, non-native English speakers working from taught phrase templates, and anyone writing to a strict rubric are over-represented in false positives. Turnitin's own documentation is candid that its indicator is a likelihood estimate rather than a determination, and it directs institutions to interpret results with human judgement [3].
The percentages deserve particular caution. Turnitin displays an asterisk (*%) instead of an exact figure when its AI detection falls below its 20% confidence threshold, which means a low-confidence signal rather than a precise measurement [3]. A document showing *% is not "clean" in any absolute sense, and a document showing 60% is not "guilty" — both are outputs of a model with a stated error rate, applied to text that may contain a mix of human and machine sentences.
Sentence-level highlighting is more informative than the headline number. A document-level score tells you how much of the text the detector considers likely AI-generated; the highlighted passages tell you which claims it is making, and those are the passages you can actually examine and defend. When a score and its highlights disagree with your own reading of the text, the disagreement is the useful information [3].
Reliability also drifts. Scores can shift when a draft is re-paraphrased, reordered, or lightly edited, and different detectors disagree with each other on the same file. Any single reading is a snapshot of one model's opinion on one version of a document, which is why re-checking after revision — and keeping the earlier reports — is more informative than chasing one number [3].
What Is the Safest Way to Check a Draft for AI Before You Submit It?
Check early, check the draft rather than the final, and check in a way that does not create a new problem for you. Running a check while you still have days to revise turns a scary result into an editing task; running it an hour before the deadline turns it into a crisis with no room to respond. The practical sequence is: finish a full draft, check it, read the highlighted passages, revise the ones that genuinely read as machine-generated, then re-check to confirm the change registered [4].
Confirm what the checker does with your file before you upload it. A check that adds your document to a student paper database can create a similarity match against your own later submission, which is a self-inflicted integrity problem no one needs. Look specifically for a non-repository check — one where the file is not added to Turnitin's student paper database and reports are not shared with third-party databases, and where you can delete your files afterwards [4]. Also check the practical constraints in advance: English-language documents, a word count above 300 and below 30,000, and a file under 20 MB cover most undergraduate and taught-master's assignments.
Keep your process evidence in parallel with the check. Save dated drafts, keep your notes and reading, and retain the reports you generate, because a report from an earlier draft plus a report from the final version tells a far more convincing story than either one alone. If a marker later raises a concern, that trail is what turns "the detector said 71%" into a documented account of how the work was written [4].
Finally, decide in advance what you will do with a bad result. If a passage is flagged and it genuinely is your own writing, the fix is usually to make it more specific — add the particular example, the named source, the awkward detail that only you would include. If a passage was drafted with a model, rewrite it in your own words rather than shuffling synonyms, and re-check. Either way, the check has done its job: it moved a hidden risk into the open while you could still act on it [4].
Reading a detector's highlights is only useful if the report you are reading is the one your institution will actually see. turnitin0 exists for that specific moment: upload your .docx, .pdf, or .txt and receive the same two PDFs a marker sees — a Turnitin AI detection report and a similarity report — in a non-repository check, typically in under 15 minutes, with files you can delete afterwards. Over 100,000 reports have been delivered to more than 20,000 students across the US, UK, Canada, Australia, New Zealand, and Ireland, and a single check is $3.80 with no subscription. You get the score, the flagged passages, and the evidence trail — before the deadline decides for you.
※ Turnitin0.com - Actual Turnitin AI Report Cover, Score, Flag And Similarity Summary
FAQ
Can a detector prove a text was written by AI?
No. Detectors produce likelihood estimates from statistical patterns, and Turnitin explicitly presents its AI indicator as a signal for human judgement rather than proof of misconduct [1]. Any decision about authorship should rest on the report plus the writing process behind it.
What does an asterisk score (*%) mean on a Turnitin AI report?
It means the detection fell below Turnitin's 20% confidence threshold, so the exact percentage is withheld and the result is a low-confidence signal [3]. It is neither a clean bill of health nor an accusation.
Why was my own writing flagged as AI?
Usually because the prose is highly predictable — formulaic transitions, uniform sentence lengths, or heavy reliance on taught academic templates. Non-native English writers and anyone writing to a rigid rubric are more exposed to this, which is why the flagged passages matter more than the overall number [3].
Does checking my draft with a third-party service count as misconduct?
A non-repository pre-submission check that does not add your file to a student paper database is a study aid, not a submission [4]. Read your institution's policy first, and avoid any service that stores your paper in a repository.
Should I rewrite flagged passages or leave them?
If the flagged text is genuinely yours, make it more specific rather than more different — add the named source, the concrete example, the detail only you would include. If a model drafted it, rewrite it in your own words and re-check, because surface synonym swaps often do not change the score [3].