Direct answer
A ChatGPT detector is a classifier that scores text for statistical patterns typical of large language models — low perplexity, uniform sentence rhythm, and predictable word choice — and returns a percentage of the submitted prose it considers AI-generated. No detector, including Turnitin's, proves authorship; it produces a probabilistic signal that a human then interprets [1]. Turnitin reports AI-writing indicators separately from similarity matches, so a clean plagiarism score tells you nothing about your AI score [1].
How does a ChatGPT detector actually decide that text is AI-generated?
Detection is a statistical comparison, not a plagiarism search. A detector does not need to find the source of your sentences; it asks whether the sequence of words looks more like the output of a language model than like writing produced by a person under normal drafting conditions. That is why a paragraph can be flagged even when it matches no existing document anywhere on the web [2].
The dominant signals are perplexity and burstiness. Perplexity measures how surprised a model is by each next word — human writing contains more unusual, less predictable choices, while ChatGPT output tends toward high-probability phrasing. Burstiness measures variation in sentence length and structure; human drafts swing between short and long sentences, whereas generated prose clusters around a narrow band. Detectors also weigh repetition, transition-word density ("moreover", "furthermore", "in conclusion"), and syntactic uniformity across a paragraph [2].
Most commercial tools now use transformer-based classifiers fine-tuned on labelled human and machine text rather than hand-coded rules. That matters practically: the classifier is only as good as its training distribution, so it performs best on the academic essay genre it was tuned for and degrades on lab reports, reflective journals, creative pieces, and heavy discipline-specific jargon [2]. Turnitin states its model was trained primarily on academic writing, which is exactly why it is used on coursework [1].
The consequence for students is that detection is sensitive to editing history, not just to who typed the first draft. A fully AI-generated paragraph that a student rewrote sentence by sentence will usually score lower than the untouched original, and a hand-written paragraph that was heavily polished by an AI tool can score higher than the author expects [2]. The detector is measuring surface texture, and surface texture changes when text passes through a model.
Why do ChatGPT detector results disagree across tools, and what does Turnitin's *% score mean?
Different detectors answer slightly different questions, so disagreement is expected rather than evidence of a broken tool. Vendors use different classifiers, different training corpora, different minimum text lengths before they will score at all, and different thresholds for what counts as "flagged". Comparing a free web checker that scores 300 words against an institutional system that analyses a 3,000-word submission is not a like-for-like comparison, and the two percentages should not be read as the same metric [3].
Turnitin's report is explicit about uncertainty. When AI detection falls below its 20% confidence threshold, the report displays *% instead of an exact percentage — a low-confidence signal rather than a precise measurement [1]. Institutional guidance stresses that the report is one piece of evidence in an academic-integrity conversation, not a standalone finding of misconduct, and that staff should look at the highlighted passages alongside drafting history and supervision records [3].
Two numbers on the same report are frequently confused. The similarity score measures overlap with existing sources in the comparison database; the AI score measures the proportion of qualifying prose the classifier attributes to AI generation. A submission can show 0% similarity and a high AI score, or 40% similarity and a low AI score. Reading the similarity percentage as a proxy for AI detection is the single most common misinterpretation students make [3].
There is also a length and formatting effect worth knowing. Detectors typically analyse prose paragraphs and may exclude quotations, reference lists, tables, and short fragments, so the denominator behind your percentage is not your whole document. A 12% AI score on a long dissertation chapter and a 12% score on a 600-word discussion post represent very different amounts of flagged text [3].
How can a student verify their Turnitin AI score before submitting?
The reliable way to know what an instructor's report will show is to run the same class of check on the same file before the deadline, rather than to rely on a free checker that uses a different model. turnitin0 produces a Turnitin AI detection report and a similarity report from one upload — the two PDFs students see in their institutional system — so the numbers you review are the numbers on the report, not an approximation from a third-party estimator [4].
Practical workflow that reduces surprises: upload the exact file you intend to submit (.docx, .pdf, or .txt, English, 300–30,000 words, under 20 MB), read the AI score and the flagged passages together, and rewrite the specific flagged sentences in your own voice rather than paraphrasing them through another AI tool. The check is non-repository — your file is not added to Turnitin's student paper database and reports are not shared with third-party databases — so a pre-submission check does not inflate your later similarity score [4].
Timing matters because revision takes longer than checking. Turnaround is under 15 minutes in 98% of cases, with most orders completing in 5–15 minutes and delivery guaranteed within 30 minutes even during queue spikes. That leaves room to act on the report instead of discovering the problem after the deadline [4]. Students who check early report the practical benefit directly: one reviewer described the report arriving "much faster than I expected" during a deadline crunch, and another noted being able to download the AI and similarity reports together and review the result [4].
Read the flagged text the way your marker will. If a passage is flagged and you can explain the drafting behind it — a source you paraphrased closely, a template you reused, a section you dictated — you are in a defensible position. If a passage is flagged and you cannot, that is the paragraph to rewrite before submission, not after [4].
Knowing how the classifier reads your text is only half the job; the other half is seeing the actual report before your marker does. turnitin0 gives you that preview — the same Turnitin AI and similarity reports your university works from — so you can decide what to revise while there is still time to revise it.
※ Turnitin0.com - Actual Turnitin AI Report Cover, Score, Flag And Similarity Summary
FAQ
Is a ChatGPT detector the same thing as a plagiarism checker?
No. A plagiarism checker compares your text against a database of existing sources and returns a similarity percentage; a ChatGPT detector classifies the statistical texture of your prose and returns an AI score. Turnitin reports both, and they are independent metrics [1].
Can a ChatGPT detector prove I used AI?
No. It produces a probabilistic signal, and Turnitin explicitly displays *% when confidence is below its 20% threshold rather than claiming precision. Institutions treat the report as one piece of evidence alongside drafting history and supervision [1][3].
Why does the same essay get different scores on different checkers?
Because each tool uses a different classifier, training corpus, minimum text length, and flagging threshold. A free checker scoring a short excerpt is not measuring the same thing as an institutional system analysing the full submission [3].
Can I check my Turnitin AI score before I submit?
Yes. turnitin0 returns a Turnitin AI detection report and a similarity report from a single upload, non-repository, with turnaround under 15 minutes in 98% of cases so you still have time to revise [4].
Does rewriting flagged passages lower the score?
Usually, if you rewrite in your own words rather than routing the text through another AI tool — which tends to restore the low-perplexity, low-burstiness patterns the classifier is looking for [2].