Direct answer
A ChatGPT detector is a tool that estimates whether a piece of text was produced by a large language model such as ChatGPT, usually by scoring how predictable the wording is rather than by proving authorship. Turnitin's AI writing detector reports the share of a document it classifies as AI-generated, highlights the matching paragraphs, and shows *% instead of a precise figure when its confidence falls below its 20% threshold [1]. The score is a probability signal a marker interprets alongside context, not a verdict, so the useful question for a student is what their own draft would show before it is submitted [3].
How Does a ChatGPT Detector Actually Work?
A ChatGPT detector does not look for a watermark or a hidden signature. It measures statistical texture: how predictable each next word is given the words before it, and how much that predictability varies across a passage. Human academic writing tends to swing between routine phrasing and unusual, topic-specific choices, while unedited model output often holds a flatter, more uniformly probable rhythm — the property detector vendors describe as low perplexity and low burstiness [2].
That is also why detection is probabilistic rather than forensic. OpenAI, which built one of the first widely used classifiers, published its own caveats: the tool was unreliable on text under roughly 1,000 characters, degraded noticeably once a passage was paraphrased or lightly edited, and produced false positives on human writing, with non-native English writers disproportionately affected [2]. A detector is therefore answering "does this read like model output?" — never "did this person use ChatGPT?" [2].
The practical consequence is that detection strength depends on the text, not on the tool's brand name. Heavily templated sections, generic literature-review prose, and formulaic transitions carry the strongest signal; original analysis, quoted data, and idiosyncratic argument carry the weakest. Two students can paste equally "AI-sounding" paragraphs and receive very different scores because the underlying predictability differs [2].
Finally, modern detectors are calibrated in bands rather than as a single yes/no switch. Because the underlying signal is a probability, vendors publish confidence thresholds and suppress low-confidence output instead of forcing a number — which is exactly why a report can show an asterisk rather than a percentage [1]. Reading a detector as a calibrated instrument, not an oracle, is the mindset that keeps you out of trouble [2].
Can Turnitin Detect ChatGPT Text, and What Does the AI Score Mean?
Yes — Turnitin's AI writing detection is designed to flag text generated by models including ChatGPT, and it reports results as a percentage of the submitted document rather than a pass/fail label [3]. The report shows an overall AI score plus paragraph-level highlighting, so a reader can see which specific passages drove the number instead of guessing from a single headline figure [3]. Turnitin's own guidance frames the output as an indicator of the possibility of AI involvement, explicitly not proof of misconduct, and expects institutions to weigh it with other evidence [3].
The score itself is the proportion of the document classified as AI-generated, expressed on a 0–100% scale, with the flagged segments shaded in the report [3]. When the detector's confidence is below its 20% confidence threshold, Turnitin displays *% instead of an exact value, and those asterisked results are low-confidence signals that should not be read as a precise measurement [1]. In practice a document showing 0% and a document showing *% are both "no strong signal," not "clean" versus "guilty" [1].
Two limits matter more than the number. First, the detector only works on English text and needs a substantial body of prose — roughly 300 words or more — before its output is meaningful, so short abstracts and reference lists are not reliable evidence either way [1]. Second, it does not identify who wrote anything; it characterises text, and a human editor who rewrites model output can move a passage out of the flagged band without changing the underlying intellectual debt [3].
Institutional practice follows from that framing. Many universities treat a high AI score as a prompt for a conversation with the student — draft history, notes, version control, supervisor discussion — rather than as an automatic penalty [3]. Knowing this changes how you should read your own report: treat the highlighted paragraphs as a list of places where your writing sounds generic, and treat the percentage as a rough index of how much of the document reads that way [1].
How Can I Check My Own Draft for ChatGPT Detection Before I Submit?
The single highest-value move is to see the signal while you can still edit, because a flagged paragraph discovered after submission can only be explained, not fixed [4]. Pre-submission checking works best as a diagnostic pass: run the draft, read the paragraph highlights, and rewrite the densest flagged blocks in your own voice with your own examples before the deadline, rather than rewriting everything indiscriminately [4].
Read the report by paragraph, not by headline number. A 30% score concentrated in two literature-review paragraphs is a very different problem from 30% scattered across your methodology, and the first is usually fixable in an hour of genuine rewriting — adding specific sources, your own data interpretation, and concrete transitions that a model would not have produced [4]. Where a passage is flagged but accurate, the fix is not deletion; it is replacing generic phrasing with the specific reasoning you actually did [4].
Keep the evidence trail alongside the score. Drafts, notes, reading annotations, and version history are what actually resolve an academic-integrity conversation, and they are far more persuasive than a screenshot of a low percentage [4]. Before any of this, confirm what your own institution permits — policies on AI assistance vary widely between departments, and a detector result means something different under a permissive policy than under a prohibition [4].
Finally, remember that a pre-submission check is a rehearsal, not an amnesty. If your draft is largely model-generated, lowering a number without changing the writing leaves the underlying problem intact, and markers increasingly ask for the process behind the text [4]. Use the report to find where your writing has drifted into generic prose, then do the writing that makes the passage yours [3].
If you want that rehearsal to use the same report format your marker will see, turnitin0 runs your .docx, .pdf, or .txt through a Turnitin check and returns the AI detection report and the similarity report together in one checkout — typically within 5–15 minutes, without adding your file to Turnitin's student paper database.
※ Turnitin0.com - Actual Turnitin AI Report Cover, Score, Flag And Similarity Summary
FAQ
Does a ChatGPT detector prove I used ChatGPT?
No. Detectors estimate how predictable the wording is and report a probability, and Turnitin describes its AI score as an indicator of possibility rather than proof of misconduct [3]. OpenAI's own classifier documentation also acknowledges false positives on human writing [2].
Why does my Turnitin AI score show *% instead of a number?
Turnitin displays *% when its confidence is below its 20% threshold, meaning the signal is too weak to report a precise percentage [1]. It should be read as "no strong AI signal," not as a hidden score.
Can I lower a ChatGPT detection score just by paraphrasing?
Light paraphrasing does reduce detector confidence, which is why classifier accuracy drops on edited text [2]. That is a limitation of detection, not a strategy: markers increasingly ask for drafts and notes, which is the evidence that actually resolves a query [4].
How much text does a detector need before the result is meaningful?
Turnitin's detection is designed for English documents of roughly 300 words or more, and short excerpts are not reliable evidence either way [1]. This is why checking a full draft is more informative than checking a single paragraph.
Is checking my draft before submitting considered misconduct?
That depends entirely on your institution's policy, which varies between departments and modules [4]. Confirming the rule first, then using a pre-submission check purely to improve your own writing, keeps the process on the right side of it [3].