Direct answer
GPT-4 output can be flagged by Turnitin's AI writing detector, but the result is a probability estimate rather than a verdict — and Turnitin itself has stated that its detector was built and trained before GPT-4-era models became widespread, so scores on newer text should be read with that limitation in mind [1][2]. The detector runs inside the Similarity Report and reports the percentage of qualifying prose it considers AI-generated, alongside sentence-level highlights that show exactly which passages drove the number [1]. Only English academic-style prose of sufficient length is assessed, and reference lists and very short strings are excluded from the calculation [1]. Because a high score can reflect formulaic academic phrasing as easily as it can reflect genuine AI drafting, the report is best treated as one signal among several rather than as proof of anything [2].
Can Turnitin's AI Detector Actually Detect GPT-4 Writing?
Turnitin's detector does produce AI-likelihood scores on GPT-4 text, but the company's own guidance frames the output as a signal that requires human interpretation rather than a determination [2]. The model underlying the indicator was developed against earlier generative output, and Turnitin has publicly acknowledged that text from newer model families — including GPT-4-era systems — may be scored differently from the text the detector was originally calibrated on [2]. In practice this cuts both ways: some GPT-4 passages are flagged confidently, while lightly edited or heavily prompted GPT-4 prose can slip under the threshold entirely.
The report structure matters as much as the headline number. Turnitin presents an overall AI-writing percentage for the submitted document plus a per-sentence view, so a reader can see whether the flagged material is concentrated in a few paragraphs or spread across the whole paper [2]. That distribution is often more informative than the total, because a single flagged block in an otherwise unflagged essay tells a very different story from uniform flagging throughout.
Turnitin also warns explicitly about false positives, particularly for writing that is highly structured, repetitive, or formulaic — exactly the register many students are taught to imitate in academic essays [2]. Non-native English writers and students working from rigid templates are frequently cited as higher-risk groups for this reason. The practical consequence is that a flagged score is not self-authenticating evidence of misconduct, and institutions are advised to weigh it against drafts, version history, and the student's own account [2].
Finally, the detector's scope is narrower than many users assume. It is designed for English academic prose, and Turnitin's own documentation describes limits on what gets analysed and how the percentage is derived from qualifying text [1]. Text outside that scope — code, equations, tables, short answers, bibliographies — is not meaningfully evaluated, so a low score on a document that is mostly non-prose proves very little [1].
What Does a Turnitin AI Score Actually Mean, and Why Does It Show an Asterisk Instead of a Number?
A Turnitin AI score is the share of qualifying text the detector classifies as AI-generated, expressed as a percentage of the document rather than of the whole submission file [3]. It is a separate metric from the similarity score: similarity measures overlap with existing sources, while the AI indicator measures stylistic likelihood of machine generation, and the two numbers can move in completely different directions on the same paper [3]. A paper can show 0% similarity and a high AI score, or high similarity and no AI signal at all.
The asterisk is the detail most students misread. When the AI signal falls below Turnitin's 20% confidence threshold, the report displays an asterisk (*) in place of an exact percentage, because the system does not have enough confidence to report a precise figure [3]. An asterisk therefore means low-confidence AI signal — not a clean bill of health and not a confirmed violation. Treating * as "I'm safe" is the single most common misreading of the report.
The sentence-level view is where the score becomes actionable. Turnitin highlights the specific segments that contributed to the percentage, so a reader can distinguish between isolated flagged sentences and sustained flagged passages [3]. If the flagged material clusters in sections you drafted with assistance, that is useful diagnostic information; if it clusters in sections you wrote entirely yourself, that is a signal the text may be tripping the detector on style alone.
Scores also need to be read against length and composition. Because the percentage is calculated over qualifying text only, a short submission or one with large non-prose components can produce a number that looks dramatic but rests on a small sample [3]. Turnitin's documentation is consistent on this point: the figure describes the detector's output, not the authorship of the student [1][3].
How Can Students Check Their Own GPT-4-Assisted Draft for a Turnitin AI Score Before Submitting?
The most reliable way to reduce uncertainty is to review a report on the draft itself rather than waiting for the institutional submission. Turnitin's own student-facing guidance encourages using report feedback on work in progress, on the reasoning that seeing how matching text and highlights are displayed turns an opaque verdict into something a student can actually revise against [4]. Reviewing early also removes the shock factor of encountering the interface for the first time on a graded submission.
Institutional access is uneven, though. Some courses enable draft or self-check submissions; many do not, and policies vary by department and by instructor [4]. Where a draft check is not available, students are left estimating their exposure from the outside — which is precisely the gap that third-party preview services exist to fill. Whatever route you take, the useful comparison is between your draft's report and the specific passages you know were AI-assisted, not between your score and a classmate's.
Revision strategy should follow the highlights, not the headline number. If flagged segments correspond to passages you generated and then lightly edited, rewriting those passages in your own voice — adding specific examples, varying sentence length, replacing generic transitions with argumentative ones — addresses the underlying cause rather than the symptom [4]. If flagged segments correspond to prose you wrote yourself, the issue is usually stylistic uniformity, and the fix is the same: introduce the irregularity that human academic writing naturally contains.
Two practical cautions are worth stating plainly. First, a clean preview does not guarantee a clean institutional report, because detector versions, thresholds, and document handling can differ between systems [2][4]. Second, a flagged preview is not a finding of misconduct — it is a prompt to revise while you still can [1][4]. Used that way, a pre-submission check is a drafting tool, not a gamble.
Reading a score is only half the job; the other half is seeing your own draft through the same lens your institution will use. That is exactly what turnitin0 is built for — an independent pre-submission check that returns a Turnitin AI detection report and a similarity report together, so you can inspect the flagged passages, compare them against the sections you know were GPT-4-assisted, and revise before the deadline rather than after it.
※ Turnitin0.com - Actual Turnitin AI Report Cover, Score, Flag And Similarity Summary
FAQ
Does Turnitin detect GPT-4 specifically, or only older AI models?
Turnitin's detector produces an AI-likelihood score on GPT-4 text, but the company has acknowledged that its model was calibrated against earlier generative output, so results on newer model families can differ from the original calibration [2]. Treat the score as an indicator, not a model-specific verdict.
What does an asterisk (*) on the Turnitin AI score mean?
An asterisk appears when the AI signal falls below Turnitin's 20% confidence threshold, meaning the system will not report a precise percentage [3]. It signals low-confidence AI signal rather than a confirmed clean result.
Is the AI score the same as the similarity score?
No. Similarity measures overlap with existing sources, while the AI indicator estimates the likelihood of machine generation; the two are reported separately and can diverge sharply on the same document [3].
Can I check my own draft before submitting it to my university?
Some institutions enable draft or self-check submissions, but availability varies by course and instructor [4]. Where it is not available, students commonly use an independent pre-submission check to review the same report format.
If my draft is flagged, does that mean I will be accused of misconduct?
Not on its own. Turnitin advises that the indicator be interpreted alongside other evidence, and its documentation notes that formulaic academic prose can trigger false positives [1][2]. A flag is best read as a prompt to revise.