Direct answer
You tell AI writing from human writing by combining two kinds of evidence: textual signals you can read yourself (flat rhythm, generic phrasing, missing personal specifics, suspiciously even paragraph structure) and a detector score from a system trained on large language model output, such as Turnitin's AI writing indicator [1]. No single signal is conclusive — Turnitin itself presents its AI percentage as an indicator that needs human interpretation, not as proof of misconduct [1]. The reliable workflow is therefore: read for signals, run a detector on the actual file, then judge the flagged passages in context.
How Can You Tell If a Piece of Writing Was Written by AI or by a Human?
Human readers pick up real signals, but they are weaker than most students assume. AI-generated prose tends to be grammatically flawless yet rhythmically flat: sentences of similar length, transitions that announce themselves ("Furthermore," "In conclusion"), and claims that stay safely general because the model has no stake in the argument [2]. Human writing usually carries friction — an oddly specific detail, an aside, a slightly awkward sentence that reveals a real person thinking. Those markers are useful for forming a hypothesis about a text, but they are exactly the markers that careful writers and editors also produce, which is why reading alone cannot settle the question [2].
Detectors work differently, and that difference matters. A system like Turnitin's analyses the statistical predictability of the text — how likely each next word is given the ones before it — because model-generated prose tends to sit in high-probability territory [1]. It then reports the share of qualifying text flagged as AI-generated rather than a single yes/no verdict [1]. That design is deliberate: the output is a probability-based indicator, and Turnitin states plainly that its score should be treated as a signal requiring contextual judgement rather than an automatic finding [1].
The practical consequence is that you should never rely on one method alone. Read the draft for the qualitative signals described above, then check the same text with a detector, then look at which passages were flagged [2]. A flag on a paragraph that you genuinely wrote from your own data is a very different situation from a flag on a section you pasted from a chatbot. The combination of signals plus flagged locations is what makes the verdict usable [2].
Finally, be honest about the limits on both sides. Human judgement is vulnerable to bias — a polished non-native speaker's essay can look "too clean," and a lightly edited AI draft can look convincingly human [2]. Detectors are vulnerable to false positives on heavily templated academic writing and to false negatives on text that has been substantially rewritten [1]. Neither side is authoritative alone, and any claim that a tool gives a 100% certain answer should be treated with suspicion [2].
What Does a Turnitin AI Detection Score Actually Mean, and How Accurate Is It?
The score is a percentage of qualifying text, not a percentage of the whole document. Turnitin's AI writing report segments a submission, excludes elements it does not evaluate (such as reference lists and very short fragments), and reports the proportion of the remaining prose that its model considers likely AI-generated [3]. This is why a 30,000-word dissertation and a 1,200-word essay with the same "40%" figure can mean very different things — the denominator is the qualifying prose, not the page count [3].
The score bands also carry information that students routinely misread. A high figure indicates a large share of the analysed text matched AI-generation patterns, while a low figure indicates little or none did [3]. Crucially, Turnitin displays an asterisk (*%) instead of an exact number when its AI detection falls below its confidence threshold — those low-confidence signals are not the same as a clean result, and they are not the same as a confirmed high score either [3]. Instructors are trained to read the report as a starting point for a conversation, and many institutions require that step before any integrity process proceeds [3].
Accuracy is the part that gets oversold online. Turnitin reports strong performance on its own benchmark evaluations, but independent researchers and universities have documented both false positives and false negatives in real submissions, particularly with heavily formulaic writing and with text that has been paraphrased or translated [3]. The honest reading is that the score is reliable enough to prompt a review and not reliable enough to stand alone as a verdict [3]. If your score is high and you did not use AI, the flagged passages themselves — not the headline number — are what you should be prepared to explain [3].
One more practical point: the score is tied to the specific text you submitted. Editing, restructuring, or rewriting flagged passages changes the input, and therefore changes the output [3]. That is why re-checking after revision is meaningful, and why a score from an earlier draft cannot be assumed to describe the version you are about to submit [3].
Can You Check Your Own Draft for AI Detection Before You Submit It?
Yes — and doing so is the single most effective way to avoid a surprise. Instructors run submissions through institutional Turnitin, which means the report exists whether or not you have seen it [4]. Checking your own draft first converts an unknown risk into a known result you can act on: you learn your AI percentage, you see which passages were flagged, and you get the similarity report in the same pass [4]. Students who check before submitting consistently describe the difference as stress reduction rather than extra work [4].
The practical requirement is that the check must use the same kind of report your institution sees. A generic "AI content detector" on a free website is not the same instrument as Turnitin's AI writing indicator, and a result from one cannot be used to predict the other [4]. What you want is a report that mirrors the institutional output — an AI percentage computed the same way, passage-level flags, and a similarity summary with matched sources highlighted [4]. Anything less leaves you guessing about the number that actually matters [4].
There is also a submission-hygiene dimension worth planning for. Institutional Turnitin typically adds submitted work to a student paper database, which is why re-submitting the same file can later register as a match against itself [4]. A pre-submission check that runs without adding your file to that repository, and that lets you delete the file afterwards, keeps your draft out of the comparison pool while still giving you the real report [4]. That distinction is easy to overlook until a similarity score comes back unexpectedly high on your own work [4].
Finally, treat the check as a diagnostic, not a certificate. If the report flags passages, the response is to revise those passages — restore your own reasoning, add the specifics only you can supply, and re-check [4]. If the report comes back clean, you have confirmation rather than a guarantee, because detection remains probabilistic [4]. Either way, you are making the decision with the same information your instructor will have, which is the entire point of checking early [4].
If you want that pre-submission verdict on your own draft — the real AI percentage, the passage-level flags, and the similarity summary in one pass — turnitin0 runs the check without adding your file to Turnitin's student paper database, and most reports come back in under 15 minutes.
※ Turnitin0.com - Actual Turnitin AI Report Cover, Score, Flag And Similarity Summary
FAQ
Is a high Turnitin AI score proof that I used AI?
No. Turnitin presents the AI percentage as an indicator that requires human interpretation, and false positives are documented on formulaic academic prose [3]. The flagged passages, not the headline number, are what an instructor should be discussing with you [3].
Can I detect AI writing without any tool, just by reading?
Partly. Flat rhythm, generic claims, and missing personal specifics are real signals, but they overlap heavily with careful human editing, so reading alone produces hypotheses rather than verdicts [2].
Why does Turnitin show a *% instead of a number?
That display indicates the AI detection fell below Turnitin's confidence threshold, so the signal is low-confidence rather than a confirmed result [3]. It is neither a clean pass nor a high score.
Does checking my draft first count as a second submission?
Not if the check runs without adding your file to Turnitin's student paper database and lets you delete it afterwards [4]. Institutional re-submission is the scenario that can register your own earlier file as a match [4].
If my draft is flagged, what should I actually change?
Revise the flagged passages rather than the whole document: restore your own reasoning, add the specifics only you can supply, and re-check the revised text, because the score is tied to the exact input you submit [3][4].