Direct answer
You check whether something was written by AI by running the text through an AI-writing detector that compares statistical patterns in the prose against models of human and machine writing, then reading the resulting report rather than a single number. Turnitin's AI writing detection, for example, reports a percentage of qualifying prose flagged as AI-generated and withholds a precise figure when its confidence falls below a 20% threshold, displaying "*%" instead [1]. No detector proves authorship on its own, so the reliable method is to combine a real report, an understanding of how the score is produced, and a review of the passages the tool actually flags [1].
How can I check if something was written by AI?
The practical starting point is to separate the three things people usually conflate: a detector, a similarity checker, and a human judgement. An AI detector looks at how the words are arranged — perplexity and burstiness, the predictability of each next token, the uniformity of sentence rhythm — while a similarity checker looks for matching sources. Turnitin reports these as two distinct outputs, an AI writing report and a similarity report, and a low similarity score tells you nothing about whether the text was machine-generated [2]. If you only run a plagiarism check, you have not answered the AI question at all.
The second point is that where you run the check changes what you learn. Institutional Turnitin access is normally mediated by an instructor or administrator, and the AI writing report is surfaced to staff through the assignment's similarity report rather than handed to students as a self-service tool [2]. That is why students looking to verify a draft before submission often use an independent pre-submission service that returns the same two PDFs — AI detection and similarity — in one checkout. On turnitin0.com, for instance, English documents between 300 and 30,000 words are processed without being added to Turnitin's student paper database, and most orders complete in 5–15 minutes.
The third point is interpretation. A detector output is a probability estimate over segments of prose, not a verdict about a person. Turnitin's own documentation frames the report as an indicator that requires human review, and the percentage refers to the share of qualifying text — meaning reference lists, quotations and very short passages may be excluded from the calculation entirely [2]. Reading the flagged passages yourself, and comparing them against your drafts and notes, is part of the check, not an optional extra.
How accurate are AI detectors, and can they be wrong?
Yes, they can be wrong — and the error is not symmetric. False positives cluster around writing that is genuinely formulaic: heavily templated lab reports, non-native English prose that has been heavily edited for fluency, and text that has passed through a grammar tool that flattened its variation. Turnitin acknowledges the limits of its model and positions the AI writing indicator as one signal among several rather than proof of misconduct [1]. Any claim that a detector is definitive should be treated as a marketing statement, not a technical one.
The confidence threshold matters more than most people realise. Turnitin does not display an exact AI percentage when its detection confidence sits below 20%, showing "*%" instead, which signals a low-confidence result rather than a clean bill of health or a confirmed hit [1]. A reader who sees "*%" and concludes "not AI" has misread the report; the correct reading is "the model could not assert a figure at its confidence bar." This is also why two detectors can disagree on the same paragraph without either malfunctioning — they are calibrated differently and trained on different data.
Detection is also inherently adversarial. As models are tuned to produce more varied, less predictable prose, the statistical gap detectors rely on narrows, and detectors are retrained in response. That race means a result is only meaningful relative to the model version and detector version in play at the time of checking [1]. For a student, the practical implication is straightforward: check the specific draft you intend to submit, in the format you intend to submit it, rather than trusting a result from an earlier version of the text.
What should I do if my text is flagged as AI-written?
First, do not rewrite in a panic. Open the report and read the flagged segments in context. Turnitin's guidance to institutions is that the indicator highlights where to look, and that the surrounding evidence — drafting history, version files, citations, supervisor conversation — carries more weight than the percentage itself [3]. Export or screenshot the report, note the flagged passages, and preserve your earlier drafts and notes, because a documented writing process is the strongest counter-evidence available to you.
Second, distinguish between the two failure modes. If the flagged text genuinely came from a model, the fix is substantive revision: rewrite the passages in your own reasoning, add specific evidence, and change the structure so the argument is yours. If the text is your own but reads as machine-like, the fix is stylistic — introduce variation in sentence length, replace generic transitions, and cut the hedging phrases that detectors associate with generated prose. Humanizing tools can accelerate the second case, but they cannot manufacture the first: meaning, citations and headings must survive the rewrite, which is why a good humanizer preserves .docx formatting rather than paraphrasing blindly.
Third, verify the fix rather than assuming it. Re-run the revised draft through the same detector you used the first time, so the comparison is like-for-like. Turnitin0 reports that 98.2% of its humanizer orders are re-checked with Turnitin for exactly this reason, and the service promises a reduction to "*%" or below 20%, or 0%, with a full refund if the score does not move for ChatGPT-, Claude- or Gemini-drafted text. Checking before and after turns an anxious guess into a measured result.
If you want that same before-and-after certainty on your own draft, turnitin0 gives you the real Turnitin AI and similarity reports — the ones your professor sees — before the deadline rather than after it.
※ Turnitin0.com - Actual Turnitin AI Report Cover, Score, Flag And Similarity Summary
FAQ
Can I check a document for AI writing for free?
Free checkers exist, but they typically return a single percentage with no segment-level detail and no similarity report alongside it. A result you cannot inspect passage by passage is hard to defend if it is ever questioned [2].
Does a 0% AI score prove the text is human-written?
No. It means the detector found no qualifying AI-generated text at its confidence level in that version of the document. Paraphrased or heavily edited AI text can pass undetected, so the score is evidence rather than proof [1].
What does "*%" mean on a Turnitin AI report?
It indicates that AI detection confidence fell below Turnitin's 20% threshold, so no exact percentage is displayed. It is a low-confidence signal, not a confirmation of either AI or human authorship [1].
Will checking my draft add it to Turnitin's student paper database?
Not with a non-repository pre-submission service. Turnitin0 checks files without adding them to the student paper database and does not share reports with third-party databases, and files can be deleted from the account afterwards.
Can my own writing be flagged as AI?
Yes. Highly templated or heavily edited prose — including non-native English writing smoothed by grammar tools — is the most common source of false positives, which is why the report should be read alongside your drafts and notes [3].