Direct answer
Whether a piece of text "is AI" is decided by statistical patterns, not by a single tell-tale phrase: detectors such as Turnitin's AI writing detector segment your prose and score how predictable and uniform the word choices are, then report the share of qualifying text that looks machine-generated [1]. Turnitin only shows an exact percentage once its confidence passes its 20% threshold; below that it displays *%, which signals a low-confidence result rather than a clean bill of health [1]. The report is evidence to be read alongside academic judgement, not a verdict on intent [1].
What makes writing look AI-generated to a detector?
Detection rests on the fact that language models pick the most probable next word far more consistently than people do. A human writer drifts — a sudden short sentence, an odd concrete detail, a comma splice at 2 a.m. — while generated prose tends to hold a steady, mid-register rhythm across hundreds of words. Turnitin's detector looks for exactly that flatness: predictable phrasing, evenly weighted sentence lengths, and transitions that arrive on schedule [2]. When a passage is flagged, the AI writing report highlights the specific segments rather than condemning the whole document, so you can see which paragraphs triggered it [2].
This is why "AI-sounding" and "AI-written" are not the same thing. A tightly edited literature review written by a careful student can score highly because academic conventions themselves reward uniformity: signposting verbs, passive constructions, hedged claims [2]. Conversely, AI text that has been genuinely reworked — new examples, specific citations, a personal argumentative line — often drops below the threshold, because the revision reintroduces the unpredictability the detector is measuring [2].
The practical takeaway is that the signals are structural, not lexical. Deleting the word "delve" or swapping "furthermore" for "also" changes almost nothing; changing how your paragraphs are built, how varied your sentence lengths are, and how specific your evidence is changes a great deal [2].
How accurate is Turnitin's AI detector, and what does an AI score actually mean?
The number in an AI writing report is a percentage of qualifying prose flagged as AI-generated, not a probability that you cheated. Turnitin calculates it only over segments long enough to analyse, which is why short submissions or heavily quoted essays can produce odd-looking results [3]. Below its confidence threshold the system refuses to commit to a figure and prints *% instead — an explicit low-confidence signal that institutions are told not to treat as an accusation [3].
Accuracy is genuinely contested at the margins. Turnitin publishes guidance acknowledging that false positives are possible, particularly for non-native English writers, highly formulaic genres, and text that has been heavily paraphrased by tools [3]. Its own advice to institutions is that the indicator should support a conversation and a review of the flagged passages, never stand alone as proof of misconduct [3]. That framing matters for students: a high score is a prompt to look at your draft, not a finding against you.
What the score is not is a measure of authorship in any philosophical sense. It cannot tell whether you outlined with a model and wrote the prose yourself, polished your own sentences with a chatbot, or generated a first draft wholesale — all three produce overlapping signals [3]. If you need to know where you actually stand, the only reliable move is to see the report on your own draft and read the highlighted segments.
Can you check whether your own draft reads as AI before you submit it?
Yes — and doing so is far safer than submitting blind. The reason students hesitate is repository exposure: a normal institutional submission can add the work to Turnitin's student paper database, which is a real concern if you are still iterating [4]. Non-repository checking exists precisely for this pre-submission stage, letting you obtain the AI and similarity reports without the file entering the paper database [4].
The workflow that actually helps is diagnostic rather than cosmetic. Run the draft, open the AI report, and read the flagged passages in order of severity: are they the sections where you leaned on a model for framing, or the sections where your own writing happens to be unusually tidy [4]? Then revise the substance — add a specific example, break a run of same-length sentences, replace a generic transition with a claim that only your argument could make — and re-check. Re-checking after revision is the step most students skip, and it is the one that tells you whether your edits actually moved the score [4].
Two practical cautions. First, check the final version, not an earlier one; formatting, headings, and reference lists all shift how text is segmented. Second, treat a *% result as "low confidence," not "cleared" — if your institution's threshold is strict, ask what number they will see rather than assuming the asterisk is neutral [4].
If you would rather see the actual report than guess from a score, turnitin0 runs the same Turnitin AI and similarity checks on your draft and returns both PDFs together, without adding your file to the student paper database. Students upload a .docx, .pdf, or .txt file over 300 words, and most reports land in under 15 minutes. It is the difference between wondering whether your writing reads as AI and reading the highlighted paragraphs yourself.
※ Turnitin0.com - Actual Turnitin AI Report Cover, Score, Flag And Similarity Summary
FAQ
Does one flagged sentence mean my essay will be treated as AI-written?
No. Turnitin highlights segments and reports the share of qualifying prose flagged, and its guidance frames the indicator as something to review rather than a finding of misconduct [3]. A single flagged paragraph is usually a prompt to look at how that passage was built.
Can a detector tell the difference between AI-written and AI-polished text?
Not reliably. Both produce overlapping statistical signals, which is why the score is a percentage of flagged prose rather than a label attached to the whole document [3]. The flagged segments are the useful part of the report, because they show you where the pattern is concentrated.
Why does my report show an asterisk instead of a percentage?
Turnitin displays *% when the AI score falls below its 20% confidence threshold, meaning the signal is too weak to report as a number [1]. Treat it as low confidence rather than as a clearance, especially if your institution applies a strict limit.
Will checking my draft add it to Turnitin's student paper database?
Not with a non-repository check, which is designed for pre-submission use so the file is not stored in the student paper database [4]. Always confirm the repository setting of whichever service you use before uploading a final draft.
What should I change if my draft is flagged?
Change structure and specificity, not vocabulary. Vary sentence length, replace generic transitions with claims tied to your own argument, and add concrete evidence — the signals detectors measure are structural, so surface-level word swaps rarely move the score [2].