Direct answer
A ChatGPT AI detector tool analyses a document and estimates how much of its prose was likely generated by a large language model such as ChatGPT. The most widely used detector in higher education is Turnitin's AI writing detection feature, which reports a sentence-level percentage of qualifying text flagged as AI-generated and is designed to be read as an indicator rather than proof of misconduct [1]. Detectors are useful for spotting patterns, but they cannot certify authorship, so results always need human interpretation [2].
How Accurate Is a ChatGPT AI Detector Tool, and Can It Tell AI Text From Human Text?
Accuracy depends on what is being measured. Turnitin publishes benchmark testing on a large corpus of human-written and AI-generated academic text, and reports a very low false-positive rate on documents that mix human and AI writing — under 1% in its own evaluation [2]. That figure matters because a false positive is the failure mode that harms students most: a fully human essay wrongly labelled as machine-written. On the detection side, the tool performs best on unedited model output, which retains the predictable rhythm and phrasing that classifiers are trained to recognise [2].
The honest limitation is that detection is probabilistic, not forensic. When AI text is paraphrased, restructured, or blended with a student's own drafting, the signal weakens, and no vendor can guarantee a correct verdict on every paragraph [2]. Turnitin itself frames the score as one data point in an academic judgement conversation rather than a standalone finding of misconduct [1]. That is why a report is most defensible when it sits alongside drafting history, version files, and citation records.
There is also a structural reason detectors miss things: they only analyse prose that meets a minimum length threshold, and they exclude reference lists, headings, and short fragments [1]. A document can therefore contain AI-assisted passages in bullet points or tables that never enter the calculation. Reading the report at sentence level — not just the headline percentage — is what tells you where the flagged stretches actually sit [1].
Finally, accuracy claims from different tools are not comparable. Vendors test on different corpora, define "AI-generated" differently, and report different metrics, so a marketing number from one checker tells you little about another [2]. For students, the practical question is narrower: does the tool match what my institution actually runs, and can I see the same report my instructor will see? That is the only comparison that changes an outcome.
Why Does Turnitin Sometimes Show an Asterisk (*%) Instead of an AI Percentage?
The asterisk is a confidence signal, not a zero. Turnitin displays *% in place of an exact figure when the AI score falls below its 20% confidence threshold, meaning the detected signal is too weak for the system to stand behind a number [3]. Students often misread this as "no AI detected" or, worse, as a system error. In practice it usually means the document contains too little qualifying AI-like prose to produce a reliable percentage.
Small or mixed documents produce the asterisk most often. Because detection runs only on prose above a minimum length, a short essay, a lab reflection, or a paper dominated by quotations and citations may leave very little analysable text [1]. A student who used ChatGPT for a few transitions and then rewrote them heavily can land in exactly this zone: some machine-like phrasing survives, but not enough to cross the threshold [3]. The asterisk is the report telling you the evidence is thin.
Asterisk results also behave differently when challenged. Because the score is below the confidence threshold, it is a weak basis for an integrity finding, and institutions are advised to weigh it with other evidence rather than treat it as a flag [3]. If you receive one, the sensible reading is "low-confidence signal — review the sentence-level highlights," not "clean" and not "caught" [1]. That distinction is worth understanding before a deadline, not after.
The practical takeaway is to look past the headline number. Open the report, check which sentences are highlighted, and ask whether those passages genuinely reflect your own drafting [3]. If the flagged text is yours and you can show your notes or version history, an asterisk is easy to explain. If the flagged text came from a model, you now know exactly which paragraphs need rewriting.
How Can I Check My Own Draft With a Turnitin AI Detector Before I Submit?
A pre-submission check works the same way an instructor's check does: the file is processed and you receive the AI writing report plus the similarity report for the same draft [4]. The value is timing. Seeing the flagged passages while you can still revise turns a stressful post-submission conversation into an ordinary editing task, and it lets you fix specific sentences instead of rewriting an entire paper [4]. Institutions increasingly expect students to be able to explain their drafting process, and a report you have already reviewed makes that explanation concrete.
What to do with the result depends on what it shows. If a paragraph is flagged and you wrote it yourself, keep the evidence — draft files, notes, earlier versions — and consider revising the phrasing that triggered the classifier, since formulaic transitions and generic topic sentences are common culprits [2]. If a paragraph came from ChatGPT and survived into your draft, rewrite it in your own words and add the analysis your marker is actually looking for [4]. The goal is not to beat a number; it is to submit work you can defend line by line.
Use a non-repository check so your draft is not added to a student paper database, and confirm the report you receive is the same format your institution uses [4]. Comparing an institutional-format AI report with a similarity report on the same file also catches the overlap problems that detectors and plagiarism scanners flag for different reasons [1]. Keep both PDFs with your version history — they are the clearest record that your check happened before submission.
One caution: repeated checking is not the same as revision. Running a draft through a detector five times without changing anything produces the same score and wastes your window before the deadline [3]. Check once, read the sentence-level highlights carefully, revise deliberately, and only re-check if you made substantial changes [4].
If you want to see the report before your instructor does, turnitin0 runs the same check on your draft and returns both PDFs — the AI writing report and the similarity report — usually in under 15 minutes, without adding your file to a student paper database. You upload a.docx,.pdf, or.txt file, and you get back exactly the document view your marker will open.
※ Turnitin0.com - Actual [Turnitin AI Report](https://www.turnitin0.com/) Cover, Score, Flag And Similarity Summary
FAQ
Is Turnitin's AI detector the same tool that checks ChatGPT text?
Yes — Turnitin's AI writing detection is built to flag text produced by ChatGPT, Claude, Gemini, and similar models, and it is the detector most universities run on submitted work [1]. Third-party checkers may use different models and produce different scores on the same file.
Can a ChatGPT AI detector prove I cheated?
No. Turnitin positions its AI score as an indicator that requires human interpretation, and institutions are advised to weigh it with drafting evidence rather than treat it as a verdict [2]. A report is a starting point for a conversation, not a finding.
What does it mean if my report shows *% instead of a number?
It means your AI score fell below Turnitin's 20% confidence threshold, so the system showed an asterisk rather than an unreliable percentage [3]. It is a low-confidence signal, not proof that the document is AI-free.
Will checking my draft add my paper to a student database?
Only if the service uses a repository submission. A non-repository check processes your file without adding it to Turnitin's student paper database, which is the option to look for when you are checking a draft before submission [4].
Does rewriting flagged sentences actually change the score?
Often, yes — especially when the flagged text is formulaic model output that you replace with your own analysis and phrasing [2]. Rewrite for substance rather than for the number, and re-check only after substantial changes.