Direct answer
GPT AI detection works by scoring the statistical predictability of your prose: detectors like Turnitin's split a document into sentence-level spans, measure how closely each span resembles the token patterns of large language models such as GPT-3 and GPT-4, and report the percentage of qualifying text that looks machine-generated [1]. The resulting AI score is an indicator, not a verdict — Turnitin displays it as *% when the overall figure falls below its 20% confidence threshold, which is why a low or asterisked score should never be read as proof of misconduct [1]. If your draft is flagged, the practical move is to understand which spans triggered it and revise those passages in your own voice before you submit.
How Does GPT AI Detection Actually Work, And What Does A Turnitin AI Score Really Measure?
Turnitin's AI writing detection is a separate report from the similarity (plagiarism) report, and the two answer different questions. Similarity compares your text against a database of published and previously submitted work; the AI indicator instead asks whether the writing itself carries the statistical fingerprint of a language model [2]. Because they are generated independently, a paper can show 0% similarity and still carry a high AI score, or the reverse — a detail students routinely misread when they only look at one number [2].
The number you see is a percentage of the document's qualifying words, not a percentage of the whole file. Turnitin excludes elements it cannot meaningfully assess, such as reference lists, quoted material, and very short fragments, then reports the share of the remaining prose that its model classifies as AI-generated [1]. That is why a 30% AI score on a 4,000-word essay does not mean 1,200 words were copied from ChatGPT; it means roughly 30% of the assessable prose matched the detector's machine-writing pattern [1].
Inside the report, flagged text is shown as sentence-level highlight bars rather than a single opaque figure. This granularity matters because it lets an instructor — or you — see exactly which paragraphs drove the score, and whether the flagged spans cluster in one section or run continuously through the whole draft [2]. Continuous, uniform flagging across an entire document is a very different signal from a handful of isolated spans in a methods section.
It is also worth being precise about what the detector is trained on. Turnitin's model was built to recognise output from GPT-3, GPT-4, and comparable large language models, and it evaluates patterns in word choice and sentence construction rather than searching for a watermark or a hidden signature [1]. There is no metadata tag inside a ChatGPT response that says "this was AI"; the signal is probabilistic and lives in the texture of the writing itself.
Why Do GPT-Written Passages Get Flagged While Human Writing Sometimes Does Not — And How Common Are False Positives?
Language models generate text by selecting the most probable next token given everything before it, which produces prose that is grammatically clean, evenly paced, and statistically unsurprising. Detectors are essentially measuring that predictability, so the traits students are often taught to value — smooth transitions, balanced sentence lengths, tidy topic sentences — are the same traits that push a span toward the machine-generated end of the scale [1]. This is the core reason a genuinely human paragraph can look "AI-like" to a model: formulaic academic register and LLM output overlap more than most people expect.
Turnitin addresses this uncertainty directly with its confidence threshold. When the overall AI score lands below 20%, the report shows an asterisk in place of an exact percentage, signalling a low-confidence result rather than a clean bill of health or a confirmed finding [3]. Crucially, an asterisked overall score can still sit alongside individually flagged spans inside the document, so the headline figure and the highlighted detail do not always tell the same story [3].
On false positives, the honest answer is that no detector publishes a per-document error rate, and Turnitin itself frames its output as a signal that requires human interpretation rather than standalone evidence of misconduct [3]. The practical implication for students is asymmetric: a high score is a reason to investigate your own draft, not a reason to assume you will be penalised, and a low score is not a guarantee that an instructor will agree with the tool [3]. Institutions are explicitly advised to weigh the report alongside drafting history, supervision conversations, and the assignment context.
The pattern that most often produces a surprising flag is mixed authorship — a human outline filled in with AI-generated paragraphs, or a hand-written essay polished sentence by sentence with ChatGPT. In those cases the detector is not wrong about the texture; it is accurately reporting that parts of the document read like model output, even though the thinking behind them was yours [1]. Understanding that distinction is what separates a fixable formatting-and-voice problem from a genuine integrity dispute.
What Should You Do If Your GPT-Assisted Draft Is Flagged Before Submission?
The first step is diagnostic, not defensive: identify exactly which spans triggered the flag rather than rewriting the entire document. Because the report highlights sentence-level evidence, you can usually trace a cluster back to a specific section — an AI-drafted literature review, a paraphrased discussion, a conclusion written in one pass [2]. Rewriting only the flagged spans preserves the work you actually did and keeps your argument intact.
The second step is to reconstruct your process in writing. Turnitin's own guidance to institutions encourages treating AI indicators as a starting point for conversation, and the strongest thing a student can bring to that conversation is evidence of how the draft was built — earlier versions, notes, outlines, dated files, supervisor feedback [4]. Keep those artefacts. A revision trail is far more persuasive than an assurance that "I wrote it myself."
The third step is to rewrite flagged passages in your own voice rather than running them through a thesaurus. Genuine revision means re-deciding what each sentence needs to say, varying sentence length deliberately, replacing generic connective phrasing with your own argumentative logic, and adding the specific examples and citations only you would choose [4]. Superficial synonym-swapping often leaves the underlying statistical pattern intact, which is why students who "edit" a flagged paragraph frequently see the score barely move.
Finally, check your draft against the same standard your institution uses before you submit it, and follow your course policy on disclosing AI assistance. Turnitin's guidance is clear that the report is meant to support academic judgement, not replace it, so the goal is a draft whose provenance you can explain confidently — not merely a number you managed to push down [4]. If a section still reads as machine-generated after revision, that is the section to rewrite from scratch rather than patch.
Knowing which spans triggered a flag is useful, but the only way to see the same AI score and sentence-level highlighting your instructor will see is to run the check yourself first — which is exactly what turnitin0 does. Upload your .docx, .pdf, or .txt and turnitin0 returns a Turnitin AI detection report and a similarity report together, generated without adding your file to Turnitin's student paper database, so you can rewrite the flagged paragraphs before the deadline instead of after it.
※ Turnitin0.com - Actual Turnitin AI Report Cover, Score, Flag And Similarity Summary
FAQ
Does Turnitin detect GPT-4 and ChatGPT text specifically?
Turnitin's AI writing model was built to recognise output from GPT-3, GPT-4, and comparable large language models, and it flags patterns in word choice and sentence construction rather than any embedded marker [1]. It does not identify which model produced a passage, only that the prose matches machine-generated patterns.
What does an asterisk instead of a percentage mean on my AI report?
An asterisk indicates the overall AI score is below Turnitin's 20% confidence threshold, so it is reported as *% rather than an exact figure [3]. It is a low-confidence signal, not proof that the document is AI-free — individual spans may still be highlighted inside the report [3].
Can a fully human-written essay get a high GPT AI detection score?
It can, particularly when the writing is highly formulaic or uses the same academic register detectors associate with model output [1]. Turnitin frames the report as a signal for human interpretation rather than standalone evidence, which is why institutions are advised to review it alongside drafting history and context [3].
Does paraphrasing flagged text with synonyms lower the AI score?
Usually not much, because synonym substitution leaves the underlying predictability of the sentences intact. Effective remediation means re-deciding what each sentence argues and rewriting it in your own voice with your own examples and citations [4].
Can I see my AI and similarity scores before I submit?
Yes. A pre-submission check returns the same two reports an instructor sees — an AI detection report and a similarity report — so you can review flagged spans and revise before the deadline, and a non-repository check keeps your file out of Turnitin's student paper database [2].