Direct answer
An AI detector tool is software that analyses a piece of text and estimates how much of it was likely produced by a generative AI model such as ChatGPT, Claude, or Gemini. Most academic detectors work at the sentence level, flagging the passages that look machine-generated and then rolling those flags up into a single document-level percentage. The number is a statistical signal, not a verdict: Turnitin's own guidance stresses that its AI writing indicator must be read alongside context and human judgement, because detection can never be certain [1]. That distinction matters, because the practical question is rarely "is this tool clever?" — it is "can I trust what this tool tells me about my draft?"
How Do AI Detector Tools Actually Work, And How Accurate Are They?
An AI detector is a classifier trained on two large piles of text: writing known to be human-authored, and writing known to be generated by a language model. The tool learns statistical regularities that separate the two — machine prose tends to be smoother, more evenly paced, and more predictable in word choice than the uneven, idiosyncratic phrasing people produce under deadline pressure. Turnitin describes its model as trained specifically to recognise the characteristics of AI-generated academic writing, then applying that model to the prose in a submitted document [2]. Crucially, the analysis is applied to prose only; reference lists, quotations, and tables are not the target of the detection model [2].
Accuracy is where marketing and reality diverge. No detector publishes a guarantee, because the underlying task is probabilistic: a detector is estimating likelihood, not recovering evidence. Turnitin's documentation is explicit that the indicator is a signal requiring interpretation, that false positives are possible, and that a flagged result is not on its own proof of misconduct [1]. In practice this means two things for anyone using a detector. First, a high score deserves a careful read of the flagged sentences rather than panic. Second, a low score is not a licence to submit something you did not write — it simply means the tool found little that matched its model of machine text [1].
The sentence-level view is the part most students overlook. Because the document percentage is an aggregate, a 30% result might mean a third of the prose is flagged throughout, or it might mean a handful of dense paragraphs carry the whole score while the rest reads clean. Turnitin segments the document into sentences and highlights the ones its model considers likely AI-generated, so the highlighted passages — not the headline number — are what you should actually read [2]. Reviewing those segments in context is the difference between using a detector as a diagnostic and using it as a slot machine [1].
What Does An AI Detection Score Actually Mean — And Why Do Some Results Show An Asterisk?
The headline percentage is best understood as "the share of the analysed prose the model considers likely AI-generated," not "the share of this essay that is AI." Turnitin's report presents the figure alongside the flagged sentences that produced it, so the score and the highlights are meant to be read as one artefact [3]. If the highlights cluster in a section you drafted with a chatbot and then edited, the score is telling you something true and useful. If they land on a paragraph you wrote from scratch in a hurry, that is the false-positive scenario the documentation warns about, and the correct response is to look at your own drafting history rather than to rewrite blindly [1].
The asterisk is the single most misread element in AI detection reports. When the detected AI writing falls below Turnitin's 20% confidence threshold, the report displays an asterisk in place of a number, because the tool does not have enough confidence to state a precise percentage [3]. Students frequently screenshot that asterisk and describe it as "0%" — which is not what it means. It means the signal is weak and low-confidence, and it is deliberately not converted into a hard figure [3]. Treating an asterisk as proof of clean text is the same category error as treating a high percentage as proof of cheating; both replace interpretation with a number the tool never claimed to give.
This is why the same draft can produce different readings on different tools, and why the report format matters as much as the score. A tool that returns only a single number gives you nothing to act on. A report that shows the document percentage, the sentence-level flags, and the similarity summary together lets you see where the signal came from and whether it overlaps with passages you were already worried about [3]. When you evaluate any AI detector tool, check what it shows you beyond the headline figure — that is the real measure of whether the output is usable.
Can You Check Your Own Draft With A Turnitin AI Detector Before You Submit It?
Yes — and doing so is a far better use of a detector than submitting blind. Turnitin's own guidance to instructors is that AI writing reports are designed to open a conversation about a piece of work, not to close one, and that flagged segments should be reviewed in context before any conclusion is drawn [4]. That logic applies just as well to the student holding the draft: if a report is going to be interpreted in context, you want to see the context before your instructor does, while you still have time to revise your own wording and explain your process [4].
The practical workflow is straightforward. Run the draft through a detector that returns a real Turnitin AI report, read the flagged sentences, and ask of each one: did I write this, or did a model write it and I left it in? Where the flagged phrasing is yours, tighten it so your own voice is visible on the page. Where it genuinely came from a chatbot, rewrite it in your own words or attribute the assistance according to your institution's policy — which is exactly the kind of revision the report is meant to prompt [4]. Reviewing flagged segments in context, rather than reacting to the aggregate number, is the behaviour Turnitin's documentation explicitly recommends [4].
Two practical cautions. First, check your institution's rules before you run anything: some universities permit pre-submission checking, others restrict which tools may be used on assessed work, and a detector result is not a substitute for reading your handbook. Second, do not let a good score become a reason to skip the read-through. A detector estimates likelihood from surface patterns; it cannot see your notes, your drafts, or your supervision meetings, and it cannot vouch for your authorship the way your own revision history can [4]. The report is a mirror, not a certificate.
If the part you actually need is not another opinion but the report itself — the document percentage, the sentence-level flags, and the similarity summary laid out the way your instructor will see them — turnitin0 runs that check on your draft and returns both PDFs together, typically in under fifteen minutes. It is an independent service, not affiliated with Turnitin, LLC, and it checks your file without adding it to Turnitin's student paper database. More than 20,000 students across the US, UK, Canada, Australia, New Zealand, and Ireland have used it to see their report before it counts.
※ Turnitin0.com - Actual Turnitin AI Report Cover, Score, Flag And Similarity Summary
FAQ
Is an AI detector tool accurate enough to rely on?
No detector is accurate enough to be treated as a verdict. Turnitin's documentation describes its AI writing indicator as a signal that requires interpretation and acknowledges that false positives are possible [1]. Use the report to find passages worth reviewing, then judge them yourself.
Why does my report show an asterisk instead of a percentage?
An asterisk appears when the detected AI writing is below Turnitin's 20% confidence threshold, meaning the tool cannot state a precise figure [3]. It signals a weak, low-confidence result — it is not the same as a confirmed 0%.
Does a high AI score mean I will be accused of misconduct?
Not automatically. Reports are intended to start a conversation, and flagged segments are meant to be reviewed in context before conclusions are drawn [4]. What matters is whether you can account for the flagged passages — through your drafts, notes, or revision history.
Can AI detectors flag text I wrote myself?
Yes. Because detection is probabilistic, human writing that happens to be smooth and predictable can be flagged, which is the false-positive case Turnitin's guidance explicitly allows for [1]. Reading the highlighted sentences is how you tell a genuine flag from a spurious one.
What should I do if my draft gets flagged?
Open the flagged sentences and check each one against your own drafting. Rewrite anything a model produced, tighten phrasing that is genuinely yours so your voice is visible, and follow your institution's policy on declaring AI assistance [4].