Direct answer
AI writing is detected by statistical signals rather than by any single "tell": detectors score text for low perplexity (predictable word choices) and low burstiness (uniform sentence rhythm), then flag the sentences that look machine-typical [1]. Provenance methods such as watermarking and metadata can help, but they only work when the generating tool cooperates [1]. No detector output is proof of authorship — the practical method is to treat a score as a review signal, confirm it against the actual draft, and check before you submit rather than after [1].
What Signals Actually Reveal AI-Generated Writing?
The most reliable signals are statistical, not stylistic. Detectors estimate how predictable each token is given the preceding text (perplexity) and how much sentence length and structure vary across a passage (burstiness); machine-drafted prose tends to score low on both because it optimises for the most probable continuation [2]. Turnitin's detector works on the same premise: it segments long-form English prose, scores only text above a minimum length threshold, and returns a percentage of qualifying text flagged as AI-generated together with sentence-level highlighting [2]. That sentence-level view matters more than the headline number, because it shows where the signal lives.
Wiki-style "tells" are much weaker than they are usually presented. Em dashes, the word "delve", triads, and tidy transition phrases appear constantly in human academic writing, and modern generators have been tuned away from them, so a checklist of vocabulary quirks produces more noise than evidence [1]. Metadata and watermarking are cleaner in principle — a signed provenance record is close to conclusive — but they require the tool to embed the mark in the first place and survive any copy-paste, retyping, or format conversion [1].
What survives all of that is document-level evidence: a version history, an outline, a reference list you can defend, and drafts that show your reasoning developing. Those artefacts are what a tutor can actually evaluate, and they are the reason detection guidance consistently frames detector output as a prompt for human review rather than a verdict [2].
Why Do AI Detectors Produce False Positives, And How Reliable Are Their Scores?
False positives are a documented, structural problem rather than a rare glitch. Because classifiers are trained on the output of specific models, their accuracy decays as generators change, and reported accuracy figures vary widely across studies and text types [1]. Formulaic writing is the main casualty: highly conventional academic prose, heavily templated lab reports, and text written by non-native English speakers all reduce the variance that detectors read as "human", so they get flagged at higher rates [1].
Score design amplifies the confusion. Turnitin reports a percentage of qualifying prose, not of the whole document, so a 40% result on a 2,000-word essay does not mean 40% of your essay was machine-written — and short passages below the minimum length are not scored at all [2]. This is also why a low-confidence signal is displayed as an asterisk rather than a precise figure: the system is telling you the evidence is weak, not that the text is clean [1].
The practical conclusion is that a detector score is one input among several. Institutions that handle this well combine the score with a conversation about sources, drafts, and the student's own understanding of the argument, because the consequences of a wrong accusation are far more serious than the consequences of a missed flag [1]. If a score looks alarming, the correct next step is to inspect the highlighted sentences against your own drafting history — not to rewrite on instinct [2].
How Can You Verify Your Own Draft Before Submitting?
Verify by reproducing the conditions your institution uses. Institutional Turnitin checks produce two separate artefacts — a similarity report and an AI writing report — and the AI report is the one that carries the percentage and sentence highlights your tutor will see [2]. Reading your draft through that same lens before submission turns an invisible risk into a fixable list of sentences.
Work from the highlighted sentences outward. For each flagged passage, ask whether you can point to the source of the phrasing: a quotation, a paraphrase with a citation, a note from your reading, or an earlier draft of your own [1]. Sentences you cannot account for are the ones worth rewriting in your own words, and rewriting them is usually faster than arguing about a score after the fact [2].
Similarity and AI are different problems and need separate fixes. A high similarity score points to matching text — usually a citation or quotation formatting issue — while a high AI score points to statistical predictability, and the remedies do not overlap [2]. Check both, because clearing one tells you nothing about the other, and a clean similarity report is not evidence that the AI report is clean [1].
Reading your own draft is useful, but it cannot tell you what a Turnitin AI report will actually return. If you want the real numbers — the same AI percentage, sentence flags, and similarity matches your tutor sees — turnitin0 runs the check before you submit, so you can fix the flagged sentences while you still have time to fix them.
※ Turnitin0.com - Actual Turnitin AI Report Cover, Score, Flag And Similarity Summary
FAQ
Does Turnitin tell you which sentences are AI-generated?
Yes. The AI writing report highlights flagged sentences alongside the overall percentage, which is why the sentence view is more actionable than the headline number [2].
Can a human-written essay be flagged as AI?
Yes. Formulaic academic prose and writing by non-native English speakers are documented higher-risk categories for false positives, because they reduce the variance detectors rely on [1].
What does an asterisk instead of an AI percentage mean?
It indicates a low-confidence signal rather than a precise score; the evidence is weak, so it should not be read as either guilt or exoneration [1].
Is a high similarity score the same as a high AI score?
No. Similarity measures matching text against existing sources, while the AI score measures statistical predictability; the two reports answer different questions and need different fixes [2].
Should a detector score decide an academic integrity case on its own?
No. Detection guidance frames scores as a prompt for human review, to be weighed with drafts, sources, and the writer's own explanation [1].