Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.

Sometimes, but no detector can reliably identify every edited or paraphrased passage. Results depend on the detector, text, language, and kind of change. Turnitin says its English AI Writing Report can flag qualifying prose it classifies as AI-generated and then modified with an AI paraphrasing tool; the company also warns that its model can be wrong. A detector score is not proof of who wrote a passage.

How can paraphrasing affect AI detection?

Paraphrasing changes wording and sometimes sentence structure, which can make AI-generated text harder for a detector to recognize. That does not mean every edit defeats every system: detection methods, text length, writing style, language, and detector updates all matter. A report generally classifies text according to patterns it detects; it does not reconstruct a writer’s editing history or establish which tool was used.

A 2023 preprint by Kalpesh Krishna, Yixiao Song, Marzena Karpinska, John Wieting, and Mohit Iyyer illustrates how much results can depend on the test setup. In one experiment, DIPPER paraphrasing reduced DetectGPT’s accuracy on text generated by GPT-2 XL from 70.3% to 4.6% at a fixed 1% false-positive rate. Those figures describe that particular model, detector, paraphraser, and experimental setting—not current performance across AI detectors or a test of Turnitin. The paper also tested a retrieval-based defense that detected 80% to 97% of paraphrased generations in its settings while classifying 1% of human-written sequences as AI-generated; that result is likewise specific to the study.

OpenAI’s educator guidance says, “In short, not in our experience,” in response to whether AI detectors work. It reports that detector research found false flags on human-written passages and that small edits can evade detection, and it cautions against using detector results for high-consequence judgments. That is OpenAI’s guidance, not an independent accuracy rating for every product.

Free tools Windows power users keep installed

One-click scans. No signup required.

Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.

Can Turnitin detect AI-paraphrased text?

Turnitin says its English AI Writing Report can classify qualifying prose as “AI-generated only” or as “AI-generated text that was AI-paraphrased.” The latter category covers text the model judges to be AI-generated and then modified with an AI paraphrasing tool or word spinner. This is a product capability claim: the classification is the model’s assessment, not proof that a particular person used a particular tool.

The feature is documented for Turnitin’s English AI detector. Turnitin says its Spanish and Japanese detectors do not include AI paraphrase and bypasser detection capabilities at present. Product support may change, so check the current model guide for the latest availability.

What the percentage means—and what it does not

Turnitin describes the overall percentage as the share of qualifying text in a submission that its model identifies as likely AI-generated, either alone or as AI-generated and then altered with an AI paraphrase tool. It is not an authorship probability and does not establish who wrote the text.

Turnitin currently marks results above 0% and below 20% with an asterisk instead of displaying a numeric score, citing a higher incidence of false positives at low scores. Reports created before July 8, 2024 may show a numeric score below 20%, so the display on an older report differs from current reports. See Turnitin’s report guide.

Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.

Which text qualifies for a report?

Turnitin’s current guidance says submissions need at least 300 words of prose in long-form writing and can contain up to 30,000 words. The report supports English, Spanish, Japanese, and Arabic, but the AI paraphrase and bypasser feature is English-only. Turnitin warns that its model does not reliably detect non-prose or unconventional formats, including poetry, scripts, code, bullet points, tables, and annotated bibliographies. A report therefore is not an assessment of every word or format in a document.

Why a detector result is not proof

Turnitin explicitly warns that its AI writing model may misidentify human-written, AI-generated, and AI-paraphrased text. It says the report should not be the sole basis for adverse action against a student. OpenAI has also said detector findings were not reliable enough for consequential student judgments. Both warnings matter because a false positive can wrongly cast suspicion on genuine human writing, while a low or zero score does not prove that writing is human-authored.

Turnitin’s October 18, 2024 whitepaper describes the vendor’s model architecture and testing protocol. Its performance statements are vendor-authored claims, not independent evidence that one accuracy rate applies to every writing task or detector. The available evidence does not establish a universal current accuracy figure for recognizing edited or paraphrased text.

Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Support on Ko-Fi

How to assess a detector claim fairly

When someone presents a detector score or compares products, check the conditions behind the result rather than treating a percentage as a verdict. Useful questions include:

Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.
  • What kind of change was tested? Light human editing, machine paraphrasing, translation, and AI bypasser tools are different conditions.
  • What text was tested? Language, genre, length, and whether the material is prose or an unconventional format can affect what a system is designed to assess.
  • Which detector and version? Models change, so a result should identify the product and version and when it was tested.
  • At what false-positive rate? Detection accuracy is meaningful only with its test population and threshold; a percentage without those details is incomplete.
  • What kind of evidence is it? Distinguish official product documentation, vendor-authored tests, and independent studies such as the 2023 preprint.
  • Who can access the report? Turnitin is an institutional product; its report should not be assumed to be available directly to every student.

What to do if a score is being used to question writing

For students and educators, treat a detector result as a prompt for review—not a conclusion about authorship. Check the applicable academic policy, then consider the writing process and surrounding evidence. Drafts, notes, sources, and version history can help explain how a piece developed. A conversation with the writer can clarify the work, but no single item automatically proves authorship or misconduct.

OpenAI’s educator guidance suggests that students may share relevant AI conversations and document how AI was used. That can provide context; it is not, by itself, proof that a particular student used an AI tool. The appropriate next step is a fair review under the institution’s policy, without relying on a detector score alone.

Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.