Direct answer
A ChatGPT detector does not look for keywords like "ChatGPT" or "AI." It measures statistical properties of your prose — chiefly how predictable each next word is (perplexity) and how much sentence rhythm and vocabulary vary (burstiness) — then reports the share of the document that reads as machine-generated. Turnitin, for example, splits a submission into roughly 300-word overlapping segments and flags the percentage of qualifying text its model classifies as AI-written, showing a *% marker when the result falls below its 20% confidence threshold [1]. A high score is a signal for a human to review, not automatic proof of misconduct [1].
How does a ChatGPT detector identify AI-generated text?
A detector starts by scoring predictability. Large language models like ChatGPT choose the most probable continuation at each step, so their output tends to sit in a narrow band of "expected" word choices. A detector's language model assigns each token a probability and aggregates those into a per-passage perplexity score; text that is consistently easy to predict looks machine-like, while human writing wanders into less expected phrasing [2].
The second signal is burstiness — variation in sentence length, clause structure, and vocabulary across a passage. Human academic writing is uneven: a long, hedged sentence followed by a short declarative one, an abrupt shift in register, an idiosyncratic word choice. ChatGPT output is smoother and more uniform, and that uniformity is measurable even when the content is perfectly correct [2].
Turnitin's own detection model was trained on a mix of human-written and AI-generated academic prose, which is why it reports results as a percentage of the document rather than a yes/no verdict. It also exposes segment-level detail, so a reviewer can see which passages drove the score instead of trusting a single headline number [2]. Because the underlying signals are probabilistic, both false positives on heavily edited human text and false negatives on lightly rewritten AI text remain possible [1].
What text patterns or artifacts make ChatGPT writing easy to flag?
The most commonly cited markers are structural rather than topical. Academic-integrity guidance points to low burstiness, near-uniform sentence length, and heavy reliance on a small set of generic connectives — "moreover," "furthermore," "it is important to note," "in conclusion" — as recurring features of unflagged LLM drafts [3]. When every paragraph opens with a topic sentence, develops three supporting points, and closes with a summary, the rhythm itself becomes a signal.
A second cluster is generic abstraction. ChatGPT tends to describe rather than specify: it names a concept but rarely supplies the granular, checkable detail — a date, a dataset, a page number, a named scholar's contested claim — that human research writing accumulates naturally. Detectors and human reviewers both react to this, and the effect compounds when citations are present but shallow [3].
A third cluster is lexical tics. Certain phrasing appears far more often in LLM output than in student writing, and studies of linguistic markers treat this over-representation as a usable feature. The practical implication is that surface-level word swaps — replacing "delve" with "explore" — rarely change a detection outcome, because the underlying predictability and uniformity remain intact [3].
Finally, artifacts can survive editing. Formatting residue, perfectly parallel bullet lists, em-dash patterns, and consistent hedging across an entire document are the kind of traces that persist after a student "cleans up" an AI draft. Reviewers are trained to look for consistency of these traces across a whole submission, since a single odd sentence proves little but a document-wide pattern is harder to explain away [1].
Can I check my own ChatGPT-assisted draft for AI detection before submitting?
In most institutions you cannot run the official check yourself. Turnitin's AI writing indicator and similarity report are delivered to instructors inside the learning management system, and the student-facing workflow does not include a self-service institutional submission [4]. That is why the first time many students see their AI percentage is after the deadline, when the report is already attached to their submission [4].
The workaround is a non-repository pre-submission check. A non-repository check processes your file without adding it to Turnitin's student paper database and without sharing the report with third-party databases, so a preview run does not itself create a similarity match against your own later submission [4]. This lets you see the AI percentage and the similarity summary on your actual draft while you can still revise.
Reading the report correctly matters as much as running it. Look at the segment-level breakdown rather than the headline figure, because a document-level percentage can be driven by a handful of passages, and pay attention to the *% indicator — it means the AI signal fell below the detector's confidence threshold, which is a low-confidence result rather than a clean bill of health [1]. If flagged passages are ones you wrote yourself, that is useful evidence to keep; if they are ones you pasted from ChatGPT, you now know exactly what to revise [2].
If you want to see the same AI percentage and similarity summary your instructor will see — on your own draft, before the deadline — turnitin0 runs a non-repository Turnitin check and returns both PDF reports, usually in under 15 minutes, so you can revise the flagged passages while there is still time.
※ Turnitin0.com - Actual [Turnitin AI Report](https://www.turnitin0.com/guides/us/turnitin-ai-checker-accuracy) Cover, Score, Flag And Similarity Summary
FAQ
Does a ChatGPT detector look for specific words?
No. Detection is driven by statistical properties such as predictability and variation in sentence structure, not by blacklisted terms, which is why simple synonym swaps usually do not change a score [2].
Why does my report show *% instead of a number?
That marker appears when the detected AI-writing percentage falls below the detector's 20% confidence threshold. It indicates a low-confidence signal rather than a confirmed zero [1].
Can ChatGPT text be detected after I edit it?
Often yes. Light editing changes wording but tends to preserve the uniformity and predictability that detectors measure, so document-wide patterns can survive a rewrite pass [3].
Is a high AI score proof that a student cheated?
No. Turnitin presents the indicator as a signal for human review, and both false positives and false negatives are possible, which is why context and drafting evidence matter [1].
Can I preview my AI score without my file entering the student paper database?
Yes. A non-repository check processes the file without adding it to Turnitin's student paper database and without sharing reports with third-party databases [4].