Direct answer
To "check AI" means running a piece of text through an AI-writing detector that estimates how much of it reads as machine-generated, and Turnitin does this by scoring only the prose segments that meet its minimum length requirement rather than the document as a whole [1]. The output is a percentage of qualifying prose flagged as AI-written, and a "*%" display means the signal fell below Turnitin's confidence threshold rather than that the text is clean [1]. Because detection is probabilistic, a single check is evidence, not a verdict — which is why understanding what the detector reads matters more than the number itself [1].
What does an AI checker actually look at when it decides whether text is AI-generated?
An AI checker does not search for a hidden watermark or a database of known chatbot outputs. It reads your prose and asks a statistical question: how predictable is each word given the words before it? Machine-generated text tends to sit in the high-probability centre of the language distribution, producing smooth, evenly paced sentences, while human academic writing carries more idiosyncratic phrasing, uneven rhythm and topic-specific vocabulary [2]. Turnitin's detector evaluates these patterns across a document and reports the proportion of prose it considers likely to be AI-generated [2].
The unit of analysis is the segment, not the essay. Turnitin only scores prose segments that contain at least 300 words of qualifying text, so short passages, reference lists, quotations, headings and tables are excluded from the calculation [1]. This is why a 250-word discussion post may produce no AI percentage at all, and why a long essay can show a score while a one-paragraph abstract does not [1]. It also explains why the percentage refers to the share of scored prose, not the share of the whole file [2].
The model is deliberately conservative at the edges. When the evidence is weak, Turnitin displays "*%" instead of a precise figure, signalling a low-confidence result below its 20% confidence threshold rather than a definitive finding [1]. Practically, this means two drafts can both show asterisks while one is entirely human and the other contains a lightly reworded AI paragraph — the tool is telling you it cannot tell [2]. Treat the score as a signal-strength indicator, not a purity test.
Finally, the detector is a language model trained on patterns, so it inherits the limits of that training. Turnitin states plainly that its AI writing indicator may not always be accurate and should be used alongside professional judgement rather than as the sole basis for an academic integrity decision [1]. For a student, the useful takeaway is that a check tells you where a reader's suspicion is likely to concentrate — the flagged segments — which is exactly the information you need to revise [2].
Why does human-written or lightly edited text sometimes get flagged as AI?
False positives are a documented limitation, not a rumour. Turnitin acknowledges that its AI writing detection can misclassify text, which is one reason the indicator is presented as a probability estimate and paired with guidance that institutions interpret it in context [3]. The detector is answering "does this read like model output?", and some genuinely human writing reads that way for entirely innocent reasons [3].
The most common cause is formulaic academic register. Literature reviews, methodology sections and definitional paragraphs are built from conventional phrasing — "numerous studies have demonstrated", "this research adopts a qualitative approach" — and that predictability is precisely the signal the detector associates with machine generation [3]. Non-native English writers are disproportionately exposed here, because heavily taught, template-driven academic phrasing can look statistically uniform even when every sentence was composed by hand [3].
A second cause is AI-assisted polishing. If you drafted by hand and then asked ChatGPT, Claude or Gemini to "improve the flow" or "make this more academic", the rewritten passages inherit the model's rhythm even though the ideas are yours [3]. Those passages are often the ones that light up, while your original paragraphs stay clean. The same applies to AI-generated outlines that you filled in manually: your sentences may be human, but the connective tissue around them is not [3].
A third cause is structural uniformity across a long document. When every paragraph opens with a topic sentence, develops three supporting points and closes with a link, the burstiness of the prose flattens — and flat, evenly paced text is what the model expects from generated writing [3]. None of this means the check is worthless; it means the flags deserve a human read. If a flagged paragraph is genuinely your own work, the fix is usually to vary sentence length and replace generic phrasing with specific, source-anchored claims [3].
How can I check my own draft for AI flags before I submit it?
The safest pre-submission check is one that does not touch the institutional paper database. Turnitin's own guidance distinguishes between a submission to an assignment (which is stored and compared against future submissions) and a draft review, and students should confirm which mode their institution uses before uploading anything [4]. A non-repository check gives you the same style of report without adding your file to the student paper database, so your own draft cannot later be matched against itself [4].
Read the report the way your instructor will. The AI writing report highlights flagged segments in the body text and reports a percentage of qualifying prose, while the similarity report separately marks matched sources — two different questions that students routinely conflate [4]. Look first at which passages are flagged, then at whether the similarity layer is showing legitimate quotations and references or genuine unattributed overlap [4]. A high similarity score from a correctly cited bibliography is not misconduct; a low AI score with one heavily flagged methods section still deserves attention [4].
Then revise by segment, not by panic. Rewrite the flagged passages in your own voice: add a specific example, name the study you are actually drawing on, break a long uniform sentence into a short one, and remove filler transitions that carry no argument [4]. If a passage was AI-polished, restore your original phrasing and keep only the edits you can defend as your own wording [4]. Re-check after revising — the point of a draft check is to iterate before the deadline, not to collect a single reassuring number [4].
Keep the report as evidence of process. Saving both the AI and similarity reports alongside your draft versions gives you a dated record of what you submitted and how it changed, which is far more persuasive in a conversation with a tutor than an assertion that you wrote it yourself [4]. Institutions increasingly ask students to discuss flagged work rather than automatically penalise it, and a documented revision trail turns that conversation into a short one [4].
If you want to see exactly what a Turnitin AI report says about your draft — the flagged segments, the percentage, and the similarity summary — before your deadline does the talking, turnitin0 runs a non-repository check and returns both PDFs together, so you can revise the specific passages that need it rather than guessing.
※ Turnitin0.com - Actual Turnitin AI Report Cover, Score, Flag And Similarity Summary
FAQ
Does "check AI" mean checking my text for AI, or using AI to check my text?
In academic contexts it almost always means the first: running your draft through an AI-writing detector to see whether it would be flagged. Turnitin's detector scores only prose segments of 300+ words, so very short texts may return no percentage at all [1].
Is a "*%" AI score good or bad?
It means the detector's confidence fell below its 20% threshold, so it withholds an exact number [1]. That is generally a low-signal result, but it is not a guarantee of originality — read the flagged segments rather than the symbol alone [2].
Why was my hand-written paragraph flagged as AI?
Formulaic academic phrasing, uniformly paced sentences and template-driven structure all resemble the statistical profile of generated text [3]. Non-native English writers and AI-polished drafts are the two groups most often affected [3].
Can I check a draft without it counting as a submission?
Yes, provided the check is non-repository, meaning your file is not added to the student paper database and cannot later be matched against itself [4]. Confirm the mode before uploading, and keep the resulting reports as a record of your revision process [4].
Does a low AI score mean my citations are safe?
No — AI detection and similarity are separate reports answering separate questions [4]. A clean AI score says nothing about unattributed overlap, so review the similarity layer for genuine matches as well [4].