Can Turnitin detect AI writing, including ChatGPT?

Direct answer: Partly, and unreliably. Turnitin's AI-writing indicator estimates how much of a document is likely AI-generated, separate from its similarity score. It flags some AI text, but also misses AI text and, more damagingly, flags human writing as AI. The output is a probability, not proof, and institutions are cautioned not to treat it as conclusive evidence alone.

Evidence: Independent testing of AI-text detection tools found that accuracy is inconsistent and degrades sharply once the text is lightly edited or paraphrased. The researchers state plainly: "the available detection tools are neither accurate nor reliable and have a main bias towards classifying the output as human-written rather than detecting AI-generated text" (Weber-Wulff et al., 2023). Detector vendors themselves report non-zero false-positive rates, which is why a flag is treated as a prompt to investigate, not a verdict.

Example: A Vietnamese Business student at the University of Manchester wrote her essay herself but had a section flagged as "likely AI." Her MAAS advisor explained the flag was a statistical estimate, not evidence, and helped her assemble her draft history to show the work was hers. The flag was an allegation she could answer, not a conviction.

Direct answer: It looks at how predictable your writing is, not whether it matches a source. AI detectors measure statistical signals, chiefly perplexity (how surprising each word choice is) and burstiness (how much sentence length and complexity vary). AI text tends to be smooth, even, and low in surprise, so writing that looks unusually uniform scores as "AI-like." Crucially, this is a correlation with AI, not a fingerprint of it: a human can produce the same pattern.

Evidence: Detection methods rely on these probabilistic language metrics rather than any record of what a tool generated, which is why every major detector reports a likelihood and warns against using the score alone (Liang et al. 2023). Because the signal is statistical, anything that makes human writing more uniform, careful editing, simple sentence structures, heavy grammar-checking, can raise the score.

Example: A Vietnamese Engineering student at UNSW wrote in short, clean, consistent sentences, the style he had been drilled in as good academic English. The detector read that uniformity as machine-like. Nothing was generated; his careful style simply matched the pattern the tool penalises, the same mechanism explained in our guide on why honest writing gets flagged.

Direct answer: Accuracy is uneven, and the errors are not random; they fall hardest on non-native English writers. Studies have shown popular detectors flag ESL writing as AI-generated far more often than native-speaker writing, because second-language writing is, on average, more uniform in vocabulary and structure. This means the students least able to argue back in fluent English are the most likely to be wrongly accused, which is precisely why a detector score cannot stand alone.

Evidence: A Stanford study found several widely used AI detectors misclassified the writing of non-native English speakers as AI-generated at strikingly high rates, while rarely misclassifying native writing, and concluded the tools are biased and unreliable as sole evidence (Liang et al. 2023). Broader testing reached the same verdict on accuracy across tools and conditions (Weber-Wulff et al. 2023).

Example: Two Vietnamese students at the same Australian university submitted equally honest essays; the one who wrote in longer, more idiomatic English passed the detector, while the one writing in the careful style she had been taught was flagged. Same integrity, different surface texture and the detector only sees texture.

Direct answer: No and MAAS will not help you do it. Tools that claim to "humanise" AI text to evade detection are unreliable, and deliberately disguising AI-generated work is an academic-integrity violation regardless of the tool used. It is a different thing entirely from defending honest work that was wrongly flagged. Evasion is also treated as an aggravating factor if discovered, turning a borderline case into a serious one. Keep the line bright: produce your own work, and there is nothing to disguise.

Evidence: Integrity policies define misconduct by the act, submitting work that is not genuinely yours, or misrepresenting how it was produced, not by which software was involved (Perkins, 2023). Because detection is unreliable in both directions, "the checker passed it" is never a defence that misconduct did not occur.

Example: A Vietnamese student once asked a MAAS advisor to "make AI text pass Turnitin." The advisor declined and explained why, then offered the real alternative: coaching him to write the piece himself, Outline → Draft → Final, so the work was genuinely his and no detector was a threat. He took that route and submitted with confidence.

Direct answer: Stay calm, do not admit to something you did not do, and gather your evidence. Collect your draft version history, outlines, notes, and annotated sources, then request a meeting and ask the institution to specify exactly what was reported and what its policy says about acting on a detector score. A documented writing process is the strongest rebuttal, because it shows the work developed over time in your hands.

Evidence: Most university misconduct procedures require corroborating evidence beyond a single detector score, and many have issued explicit cautions that AI-detection results are not reliable as sole proof. Your process record, timestamps, edits, comments, is the recognised safeguard precisely because the detector cannot see it.

Example: A Vietnamese postgraduate at the University of Glasgow felt his English was not strong enough to argue back after an AI flag. His MAAS advisor helped him present his document version history showing weeks of incremental edits, plus his handwritten outline. Faced with the drafting trail, the panel cleared him, the process record did the talking.

Direct answer: Write the work yourself and keep evidence that you did. Draft in a tool that preserves version history, save your outlines and notes, let your natural voice vary in sentence length and word choice, and use AI, if your institution permits it, only in ways you could openly disclose. None of this is about gaming the detector; it is about leaving an honest, visible trail so that a flag, if it ever comes, is easy to answer.

Evidence: Because detectors are unreliable in both directions, neither evading nor satisfying them is a sound strategy; the durable safeguard is genuine authorship plus a documented process, which is also what a viva or an integrity panel ultimately tests (Weber-Wulff et al. 2023). Authentic human writing naturally varies in rhythm, the opposite of the flat uniformity detectors penalise.

Example: After one scare, a Vietnamese Marketing student at RMIT adopted her MAAS advisor's habit: draft only in a version-tracked document, keep every outline, and read each paragraph aloud so her own voice came through. Her later submissions carried a clear authorship trail, and she stopped fearing the submission button.