Direct answer
Detecting AI-generated text is a statistical exercise, not a plagiarism match: detectors look for patterns in how words are chosen and sequenced, then estimate the probability that a machine produced a given passage. Turnitin's AI writing detector works this way, scoring submitted prose and reporting the share of qualifying text it believes was generated by an AI tool [1]. Because the output is a probability rather than a proof, the number you see needs interpretation before you act on it. This guide explains what detectors actually measure, where they fail, and how to test your own draft before it reaches a marker.
How Do AI Detectors Actually Tell Whether Text Was Written by AI?
AI detectors do not search a database of "AI sentences" the way similarity checkers search a paper repository. Instead they model language statistically and ask how surprising each next word is, given everything before it. Human academic prose tends to be uneven: a writer reaches for an unusual verb, repeats a phrase, or takes a syntactically awkward turn. Language models optimise for the most probable continuation, which produces text that is fluent, evenly paced, and predictable at the token level. That predictability gap is the core signal [2].
Turnitin's approach layers that statistical model over structural checks. The system segments a submission into sentence-level ranges, evaluates each range, and only scores segments long enough to give a reliable signal, which is why short answers or bullet lists may return little or no AI indication [1]. The report then aggregates flagged ranges into an overall percentage of qualifying text. This is why a document can show a moderate score while individual highlighted passages look scattered rather than continuous — the score describes proportion of text, not the writer's intent [1].
Two further signals matter in practice. Detectors weigh burstiness — the variation in sentence length and complexity — so a draft where every sentence runs 18 to 22 words reads as machine-regular. They also weigh lexical diversity and repeated transitional scaffolding such as "moreover," "furthermore," and "it is important to note," which models over-produce relative to human writers [2]. None of these signals is evidence on its own; detection is the combined weight of many weak signals.
The practical consequence is that detection is probabilistic and threshold-based. Turnitin reports a *% instead of an exact figure when its AI detection confidence falls below its 20% threshold, meaning the signal is weak and should be treated as low-confidence rather than as an accusation [1]. Reading the report correctly — score, flagged ranges, and confidence display together — tells you far more than the headline percentage alone.
How Accurate Are AI Detectors, and Why Do They Sometimes Flag Human Writing?
Accuracy is the question every reader of a detection report should ask first, and the honest answer is that no detector is uniformly accurate across genres, languages, and writing styles. Detectors are trained and validated on particular corpora, so their performance is strongest on the kind of academic prose they were tuned for and weaker on highly technical writing, non-native English, and formulaic genres such as literature reviews [3]. A detector's output is a likelihood estimate calibrated to its training data, not a verdict about authorship.
False positives are a documented limitation rather than a rare glitch. Highly structured human writing — lab reports, methods sections, exam-style essays taught to a template, and text written by multilingual students using careful, conventional phrasing — can look statistically regular enough to trigger flags [3]. So can heavily edited prose, where a writer has smoothed away idiosyncrasies in revision. Turnitin itself frames its AI writing indicator as one data point for educators to weigh alongside drafts, version history, and conversation, not as standalone proof of misconduct [3].
There is also an asymmetry worth understanding. Detectors tend to be more reliable at identifying fully AI-generated text than at separating AI-assisted from human writing, because light editing, paraphrasing, and humanising rewrites deliberately reduce the statistical signature the detector relies on [3]. That cuts both ways: a low score does not certify that no AI was used, and a high score does not certify that AI wrote the passage.
For students, the defensible response is process evidence plus a pre-submission check. Keep drafts, version history, notes, and reference trails — the material that shows how the argument developed — and treat any detection score as a signal to inspect, not a result to panic over [3]. Where a flag lands on a passage you wrote yourself, being able to show the drafting trail is far more persuasive than arguing about the percentage.
Can You Check Your Own Draft for AI Detection Before You Submit It?
Yes — checking a draft before submission is now a normal part of academic workflow, and it is the single most useful step for reducing surprise. Turnitin's guidance for students positions pre-submission checking as formative feedback: the point is to see which passages a detector reads as machine-like while you can still revise them [4]. Institutional access varies, so the workflow you can use depends on whether your university licenses a draft-checking tool or you check independently.
When you do check, replicate the conditions of the real submission as closely as possible. Upload the same file format you will submit, at full length, because detectors score only qualifying prose segments and a truncated or reformatted file changes the ranges they evaluate [4]. Read the report the way a marker would: locate the flagged ranges, then look at what is actually happening in those sentences — uniform sentence rhythm, stacked transitions, generic hedging, or a list of abstract nouns with no concrete referent. Those are the passages to rewrite in your own voice, with specific examples, citations, and varied sentence length [4].
Interpret the number with the confidence display in mind. A *% result means the detector's confidence sat below its threshold, so the finding is weak rather than conclusive [4]. A high percentage concentrated in a few paragraphs is a targeted revision task; a high percentage spread evenly across the whole document usually points to a drafting process problem rather than a handful of sentences, and that is worth addressing at the outline-and-notes stage next time.
Finally, keep the report. Saving the pre-submission report alongside your drafts gives you a dated record of what the text looked like before you revised, which is exactly the evidence that resolves most misunderstandings [4]. Checking is not about chasing 0%; it is about knowing where your prose reads as machine-regular and fixing it while the fix is still cheap.
Reading a detection score is only useful if you can read it on your own draft, at full length, before the deadline does the reading for you. That is what turnitin0 was built for: upload your.docx,.pdf, or.txt and receive the Turnitin AI detection report and similarity report together, showing the score, the flagged passages, and the match highlights exactly as a marker would see them — typically back in under 15 minutes, without your file entering Turnitin's student paper database. Over 100,000 reports delivered to more than 20,000 students, rated 4.9/5.0.
※ Turnitin0.com - Actual Turnitin AI Report Cover, Score, Flag And Similarity Summary
FAQ
What does a *% AI score mean on a Turnitin report?
A *% display means Turnitin's AI detection confidence fell below its 20% threshold, so the signal is low-confidence rather than a precise measurement [1]. Treat it as a weak indication and inspect the flagged ranges rather than reading it as a fixed percentage.
Can AI detectors prove that a student used ChatGPT?
No. Detectors produce probability estimates about text patterns, and Turnitin presents its AI writing indicator as one data point to be weighed with drafts, version history, and other evidence [3]. A score is a prompt for review, not proof of authorship.
Why does my own writing get flagged as AI-generated?
Formulaic or highly structured prose, heavily revised text, and careful conventional phrasing from multilingual writers can all read as statistically regular to a detector [3]. Rewriting flagged passages with specific detail and varied sentence length usually reduces the signal.
Do paraphrasing tools prevent AI detection?
Light paraphrasing and humanising rewrites reduce the statistical signature detectors rely on, which is why separating AI-assisted from human writing is harder than detecting fully AI-generated text [3]. That does not make the underlying process question go away — and a low score never certifies that no AI was used.
Should I check my draft before submitting it?
Yes. Pre-submission checking lets you see which passages read as machine-like while you can still revise them, and saving the report alongside your drafts gives you a dated record of the text before revision [4]. Upload the same file format and full length you intend to submit.