Direct answer
An AI detector for GPT text does not search for copied sentences; it analyses statistical patterns in your writing — how predictable each next word is and how much sentence rhythm varies — and flags passages that look machine-generated. Turnitin's AI writing report then shows the percentage of qualifying text it attributes to AI, broken down by segment [1]. GPT-family output is the main target of these detectors, but no detector identifies which model wrote a passage, and no score is a verdict on its own [2].
How Does an AI Detector Tell Whether Text Was Written by GPT?
Detectors such as Turnitin's evaluate how a sentence is built, not what it says. The model scores text for predictability: human writing tends to contain unexpected word choices, uneven clause lengths, and small stylistic irregularities, while GPT output is optimised to pick the most probable next token, which produces unusually smooth, evenly weighted prose [2]. Turnitin describes its detector as identifying AI-generated writing including output from GPT-family models, and it does not attempt to name the specific model used [2].
The practical consequence is that detection is probabilistic, not forensic. A detector outputs a likelihood over a stretch of text, which is why the report is expressed as a percentage of qualifying text rather than a binary "AI / not AI" label [1]. Turnitin's own guidance treats the score as one signal to be read in context, not as proof of misconduct [2]. Two students can submit equally original work and receive different percentages simply because one writes in a more regular, formulaic register.
It also matters which text counts. Turnitin's report only evaluates prose-like, qualifying text, and excludes elements such as quotations, references, and very short fragments from the percentage [1]. A long reference list or a block of direct quotation can therefore make a report look smaller than the body of the essay suggests — and a heavily bulleted or tabular submission may produce very little qualifying text at all.
Why Does GPT-Written Text Sometimes Score 0% or *% on the AI Detector?
A *% result is not a clean bill of health. Turnitin displays an asterisk instead of a number when the amount of detected AI writing falls below its 20% confidence threshold, meaning the signal is too weak or too small to report as a confident percentage [3]. The same display can appear on a short submission, on text that has been substantially rewritten or paraphrased, or on a draft where only a small fraction of sentences are machine-generated [3]. Treat *% as "low-confidence," not "verified human."
The reverse surprise is just as common: genuinely human writing can be flagged. Formulaic academic registers — literature-review summaries, methods descriptions, definitional opening paragraphs — share the very features detectors associate with GPT output: predictable transitions, balanced sentence structure, and low lexical surprise [2]. Students who write in a heavily templated style, or who use grammar tools that flatten their phrasing, are more exposed to this than students with idiosyncratic prose.
Editing history changes the result too. When GPT output is manually adjusted, sentence-level rewrites break up the uniform rhythm the detector keys on, so a draft that scored high in its raw form may drop sharply after revision — or may drop in some paragraphs while others stay flagged [1]. Because the report is segmented, a single flagged section can sit inside an otherwise clean essay [1], which is why reading the segment breakdown matters more than reading the headline number.
Can You Check a GPT-Assisted Draft With a Real Turnitin AI Report Before You Submit?
Yes — and doing so is the single most useful step before submission. Turnitin's own guidance to institutions is that the AI score should be weighed alongside drafting history, revision evidence, and a conversation with the student rather than acted on in isolation [4]. That logic applies to you as the author too: if you can see the same report your instructor will see, you can judge whether a flagged segment reflects a real problem or a false positive, and you can keep the drafts and revision history that support your authorship [4].
Students who understand what the report measures are also less likely to trip it accidentally. Turnitin notes that discussing AI writing openly with students — what the indicator means, what it does not mean, and how drafting records help — reduces both accidental flags and unnecessary panic [4]. In practice that means keeping earlier versions of your document, noting where you used an AI assistant for brainstorming or phrasing, and checking the finished draft rather than assuming a low-effort final pass removed the signal.
The catch is access: the report lives inside an institutional LMS, and most students cannot generate one on demand. That is exactly the gap turnitin0 fills. You upload a .docx, .pdf, or .txt file (English, 300–30,000 words, under 20 MB) and receive two PDFs in one checkout — a Turnitin AI detection report and a similarity report — matching what professors see, typically in under 15 minutes, with the file checked non-repository so it is never added to Turnitin's student paper database. Over 100,000 reports have been delivered to 20,000+ students, and a single check is $3.80.
If your GPT-assisted draft comes back with a flagged segment, the useful next move is to see the actual report rather than guess at the number — turnitin0 shows you the same cover, score, segment flags, and similarity summary your instructor will open, so you know precisely which paragraphs need rewriting before the deadline.
※ Turnitin0.com - Actual Turnitin AI Report Cover, Score, Flag And Similarity Summary
FAQ
Does an AI detector know which GPT model wrote my text?
No. Turnitin's detector identifies AI-generated writing including GPT-family output but does not attribute text to a specific model [2]. The report tells you that a segment looks machine-generated, not whether it came from GPT, Claude, or Gemini.
Is a 0% AI score proof that my writing is human?
Not proof — just a low signal. Turnitin shows *% when detected AI writing falls below its 20% confidence threshold, which means the evidence is weak or limited rather than absent [3]. Heavily edited or short text can land there even when AI assistance was used.
Why is only part of my essay flagged?
The report evaluates qualifying prose and reports results by segment, excluding quotations, references, and very short fragments [1]. A flagged paragraph usually reflects a stretch of uniformly predictable phrasing rather than the whole document.
Can I lower the score by paraphrasing the flagged paragraphs?
Sometimes, because rewriting disrupts the uniform rhythm detectors respond to [1]. But paraphrasing alone is unreliable and can damage meaning or citations; rewriting flagged passages in your own voice, and keeping earlier drafts as evidence of authorship, is the safer approach [4].
What do instructors actually do with the AI score?
Turnitin advises treating the indicator as one piece of evidence alongside drafting history and discussion with the student, not as a standalone finding [4]. That is why keeping your revision trail is as important as the number itself.