Direct answer
To check whether a text was written by AI, run it through a detector that reports a percentage of qualifying text rather than a yes/no verdict, then read that percentage alongside the flagged segments, the length of the text, and whether it was edited after generation. Turnitin's AI writing report is the reference standard for academic submissions because it is the same report instructors see, and it separates AI-generated from AI-paraphrased content [1]. No single score proves authorship, so the reliable method combines a detector result with structural signals and human judgement.
How Do AI Detectors Actually Decide If Text Was Written By AI?
AI detectors do not recognise "AI words." They score statistical patterns that are more probable in machine-generated prose than in human prose — token predictability, sentence-length uniformity, and low lexical variance across a passage. Turnitin's detector runs a model trained on large corpora of both human and AI writing and returns a percentage of the document's qualifying text rather than a binary flag [2]. That design choice matters: the output tells you how much of the text looks machine-like, not whether a person was involved in producing it.
The unit of analysis is the segment, not the document. A report can show an overall percentage while also highlighting specific sentences as AI-generated, which lets a reader see whether the signal is concentrated in one paragraph or spread evenly through the whole piece [1]. Very short texts are a known limitation — there is not enough signal in a 60-word paragraph for the model to work with, so short submissions produce unstable results [2].
Two further mechanics shape what you see. First, Turnitin distinguishes AI-generated text from AI-paraphrased text, because a passage rewritten by a tool leaves a different statistical footprint than one drafted from scratch [2]. Second, low-confidence detections are displayed differently rather than as a precise number, which is a deliberate signal that the evidence is weak [2]. Reading the report without knowing these conventions is the most common reason students misjudge their own results.
Practically, this means you should never treat a detector as a lie detector. It is a classifier producing an estimate under uncertainty, and the estimate is only as good as the text you feed it [1]. The correct question is not "did AI write this?" but "how much of this text behaves like AI output, and where?"
How Accurate Are AI Detectors, And What Do The Results Really Mean?
Accuracy is highly dependent on the text, and the published evidence is consistent on this point: detection performance falls sharply once a text has been paraphrased or lightly edited by a human. A widely cited evaluation found that detector accuracy dropped substantially on machine-generated text that had been run through paraphrasing tools, and that detectors struggled to separate heavily edited AI text from genuine human writing [3]. In other words, the more a student edits a draft, the less reliable the score becomes in either direction.
The fairness problem is equally documented. The same research line shows false positives clustering around writing by non-native English speakers, whose prose can share surface features — simpler vocabulary, more formulaic transitions — with machine output [3]. This is why no institution should treat a detector percentage as standalone proof, and why Turnitin itself frames the report as an indicator that requires academic judgement rather than an automatic penalty trigger [4]. If you are the one being assessed, that framing is your strongest argument: the report is evidence to be interpreted, not a verdict.
What the numbers mean in practice is narrower than most people assume. A high percentage means a large share of the qualifying text carries machine-like statistical patterns; it does not identify which tool was used, when, or with what intent [1]. A low percentage means the text did not trip the model's thresholds — it is not a certificate of human authorship, especially for short or heavily edited passages [2]. And a mid-range score is the least informative outcome of all, which is exactly when a second check on a longer, unedited sample is worth the effort.
The practical takeaway is to interpret results in context. Compare the score against how the text was actually produced, check whether the flagged segments correspond to passages you know were drafted with assistance, and treat any single number as one input among several [3].
How Can You Check A Text With Turnitin Before You Submit It?
The most reliable pre-submission check is to run the text through the same detection engine your institution uses, because a different detector's score tells you nothing about what your professor will see. In the normal Turnitin workflow, the AI writing report is generated inside the assignment and returned to the instructor, which is why students usually cannot retrieve their own report after submitting [4]. Previewing before submission closes that gap: you see the AI percentage and the similarity summary that will appear in the marking workflow, while you can still revise.
A sound checking routine has three steps. First, confirm the text is long enough to be scored meaningfully — short extracts produce unstable results, so test a full section rather than a single paragraph [2]. Second, read the flagged segments rather than only the headline percentage, since a concentrated flag in one paragraph is a very different problem from a diffuse one across the document [1]. Third, re-check after any revision, because paraphrasing changes the statistical profile and can move a score in either direction [3].
It also helps to know what the report will not tell you. It will not state which model produced the text, and it will not resolve intent; those remain academic judgements that your institution is expected to make using the report as one input [4]. Treating the preview as diagnostic rather than judicial keeps the exercise useful — you are looking for passages that need rewriting, not for a clean bill of health.
For students in the US, UK, Canada, Australia, New Zealand, and Ireland working to tight deadlines, the practical constraint is turnaround. A preview is only useful if it arrives before the deadline does, so a check that returns in minutes rather than days is what makes the revise-and-recheck loop realistic.
If you would rather see the actual report than guess from a percentage, turnitin0 runs your draft through Turnitin's detection engine and returns the AI and similarity PDFs together, so you know exactly what is being flagged before your professor does — no subscription, and the file is never added to the student paper database.
※ Turnitin0.com - Actual Turnitin AI Report Cover, Score, Flag And Similarity Summary
FAQ
Can I check a text for AI without submitting it to my university?
Yes. A pre-submission check runs the document through the same detection engine outside the assignment workflow, so no report is delivered to your instructor and the file is not added to the student paper database.
Does a 0% AI score prove a human wrote the text?
No. A low score means the text did not cross the detector's thresholds, which is weaker evidence for short or heavily edited passages than most students assume [2]. It should be read as one signal, not a certificate.
Why does my score change after I edit the text?
Paraphrasing and restructuring alter the statistical patterns the detector measures, and published evaluations show accuracy shifts noticeably on edited text [3]. Re-checking after revision is the only way to know where you stand.
What is the difference between AI-generated and AI-paraphrased in the report?
Turnitin flags the two separately because a passage rewritten by a tool leaves a different statistical footprint than one drafted entirely by a model [2]. Both categories count toward the overall percentage but mean different things about how the text was produced.
How long should the text be for a meaningful check?
Long enough to give the model a stable sample — a full section or complete draft rather than a single paragraph, since very short texts produce unreliable results [2].