Direct answer
Detecting AI-generated content online means looking for statistical fingerprints rather than "tells": modern detectors score how predictable each word is given the words before it, then report the share of text that looks machine-like [1]. No single signal — not em dashes, not a "delve" or "tapestry," not a gut feeling — is reliable on its own, which is why serious detection combines close reading, metadata checks, and a calibrated detector report [3]. Turnitin's AI writing indicator, for example, reports the percentage of qualifying prose it flags and shows an asterisk instead of a number when its confidence is low [1]. Treat any single score as evidence to weigh, never as a verdict.
What are the most reliable ways to tell if text was written by AI?
The most reliable approach is layered: read for substance and specificity, check provenance and metadata, then run a detector whose scoring method you actually understand. Turnitin's AI writing detection works by segmenting a document into sentence-level spans and estimating, for each span, how statistically predictable the word choices are [1]. It reports the proportion of qualifying text flagged, and it deliberately withholds a precise figure when the signal is weak — a display of *% means the result sits below the confidence threshold rather than meaning "0% AI" [2]. That design choice matters: the tool is built to be read by a human who then investigates, not to output a pass/fail stamp [2].
Close reading still earns its place. Generic hedging, symmetrical paragraph lengths, and paragraphs that restate the prompt before answering are weak signals individually but add up. Stronger signals are factual: citations that don't resolve, quotes that don't exist, statistics with no traceable source, and a bibliography whose sources are all plausible-sounding but unverifiable. Those are verifiable in minutes and don't depend on any detector's accuracy [3].
Provenance checks are the most underrated layer. Ask for the working file, not the PDF — version history, comment threads, and revision timestamps in Google Docs or Word are hard to fabricate convincingly. For published web content, look at the author's footprint: a byline with no other writing, a domain registered weeks ago, and images with AI-generation metadata all point the same direction. None of these is proof alone, but they compound.
Finally, match the tool to the job. A detector tuned for long-form academic prose behaves differently from one built for social posts or code. Turnitin's indicator, for instance, only reports on documents with enough qualifying prose to produce a defensible percentage [2], so a 150-word paragraph is the wrong input for it.
How accurate are AI detectors, and why do they sometimes flag human writing?
Detectors estimate the probability that a model would have produced a given sequence of words, so anything that makes human writing statistically predictable pushes its score up [3]. Formulaic genres are the classic case: lab reports, literature reviews, and exam-style essays follow conventions so tightly that a careful human writer can look machine-like. Writing produced by someone working in a second language can also score higher, because a smaller, more repetitive vocabulary is exactly what the detector associates with generated text [3].
This is why false positives are a documented, acknowledged limitation rather than a bug in one vendor's product. Turnitin publishes guidance on its false-positive rate and frames the AI writing indicator as one data point requiring human judgement, not as an accusation [3]. The practical implication for anyone detecting content is to weight results by document length and genre: a 40% flag on a 3,000-word original analysis means something different from a 40% flag on a 400-word discussion post that follows a template.
Accuracy also degrades in predictable ways. Heavily edited AI text — paraphrased sentence by sentence, with new examples inserted — is the hardest category, because the statistical signature is genuinely diluted. Short texts, poetry, tables, and reference lists are outside what most detectors can score meaningfully [2]. And detectors are not stable over time: model updates change the underlying distributions, so a threshold that worked last year may not hold this year.
The honest conclusion is that detection is probabilistic evidence, and its value depends on how you use it. Use a high score to decide where to look, not what to conclude. Use a low score to rule out the crudest forms of copy-paste generation, not to certify authorship. And always let the writer respond — a five-minute conversation about sources resolves more cases than any percentage [3].
How can you check a document against Turnitin's AI detector before you submit it?
Turnitin's AI writing report is the same view your instructor sees: an overall percentage plus highlighted passages, so you can see exactly which sentences carried the signal [4]. The catch is access. Students can only see that report after a submission has been made through an institutional account, and whether they can see it at all depends on how the assignment was configured [4]. There is no official student-facing "check my draft" button inside Turnitin itself.
That gap is why pre-submission checking services exist. Turnitin0 runs your .docx, .pdf, or .txt file through a non-repository check and returns two PDFs in one order — an AI detection report and a similarity report — formatted the way an instructor's LMS displays them. Because the file isn't added to Turnitin's student paper database, checking a draft doesn't create a match against itself later. Documents need to be English, over 300 and under 30,000 words, and under 20 MB; most orders return in under 15 minutes, with delivery guaranteed within 30 minutes even during queue spikes.
Reading the report is the part most people skip. Start with the overall AI percentage, then open the flagged spans and read them in context — a flagged methods section that paraphrases a standard protocol is a very different finding from a flagged original argument. Check the similarity report separately, because a high similarity score and a high AI score are different problems with different fixes. If the AI score is a *%, you are looking at a low-confidence signal below the 20% threshold, not a clean zero.
Then act on what you find. If specific passages are flagged and they're genuinely yours, rewrite them with your own examples and data — specificity is the single most effective de-optimisation. If the flagged text came from a model and you need it to read as your own work, rewriting by hand is slow but reliable; a humanizer is faster, and Turnitin0's service targets *% or 0% on ChatGPT, Claude, and Gemini output with a full refund if it misses.
Knowing what a detector sees is only useful if you can see it too — and the moment before submission is when that matters most. Turnitin0 gives you the actual report, not a guess: real Turnitin AI and similarity PDFs you can read line by line, with flagged passages highlighted exactly as your instructor will see them, delivered in minutes and without adding your draft to any database.
※ Turnitin0.com - Actual Turnitin AI Report Cover, Score, Flag And Similarity Summary
FAQ
Does Turnitin tell you how it detects AI writing?
At a high level, yes: it scores how predictable each sentence-level span is and reports the share of qualifying text that looks machine-generated [1]. Turnitin does not publish its model weights or exact thresholds, which is normal for detection vendors.
What does an asterisk in the AI score mean?
It means the result fell below Turnitin's confidence threshold — the tool is signalling a weak, low-confidence detection rather than a clean zero [2]. Read *% as "not enough signal to report a number," not as "definitely human."
Can AI detectors prove someone cheated?
No. Detector output is probabilistic evidence, and vendors including Turnitin document false positives and expect human review before any conclusion is drawn [3]. A percentage should start an investigation, not end one.
Why does my own writing get flagged as AI?
Formulaic structure, heavy use of stock transitions, and a narrow vocabulary can all raise a score, and writing in a second language is a recognised contributor [3]. Adding specific data, named sources, and your own examples lowers the statistical predictability that triggers flags.
Can I see my AI report before I submit?
Only if your institution's assignment is configured to let you, and only after submission [4]. Pre-submission services like Turnitin0 return the same report format on a draft without adding it to Turnitin's student paper database.