AI paraphrasing tools — sometimes called “text spinners,” rewording tools, or paraphrasing generators — take a block of text and produce a reworded version that keeps the same underlying ideas and structure while changing the words and sentence patterns. That rewording can reduce the textual overlap a similarity checker like Turnitin or iThenticate reports against a source, even though the substance of the passage, and the fact that it originated with someone else, has not changed. This guide is written for research-administration and academic-integrity offices: what these tools actually do to a similarity score, why a low score is not the same thing as original work, how this differs from legitimate AI-assisted language editing, and how institutions and detection vendors have adapted their policies and tooling in response.
How similarity checkers work, and why paraphrasing defeats them
Similarity-detection systems — Turnitin, iThenticate, and Crossref Similarity Check among the most widely used in academic and publishing contexts — work by comparing submitted text against a large reference corpus (previously submitted papers, journal archives, open web content) and flagging overlapping strings of text, typically matched at the level of short word sequences. See CASRAI’s guide to how Turnitin, iThenticate, and Crossref Similarity Check work for the underlying mechanics. That approach is very effective at catching verbatim or lightly-edited copying, but it depends on the submitted wording resembling the source wording closely enough to register as a match.
An AI paraphrasing tool breaks that assumption directly. By substituting synonyms, restructuring clause order, changing sentence voice, and varying phrase length, it can produce output that shares essentially no exact word sequences with the source text while preserving the source’s argument, evidence, and organization word-for-word in substance. The similarity score the tool reports can drop sharply as a result — not because the writing became more original, but because the detection method’s core assumption (verbatim or near-verbatim overlap) no longer holds.
This is not a hypothetical evasion technique. It overlaps with a pattern research-integrity investigators already track in a different but related context: paper mills and some questionable-research operations have used automated synonym substitution to disguise recycled or fabricated text, producing the “tortured phrases” (odd substitutions like “counterfeit consciousness” for “artificial intelligence”) that CASRAI’s guide to paper mills and tortured phrases covers as a research-integrity red flag. AI paraphrasing tools are a more fluent, harder-to-spot version of the same underlying evasion logic.
Is AI-paraphrased text plagiarism?
The similarity score a detector reports is evidence, not a definition. Whether a passage constitutes plagiarism turns on whether someone else’s ideas, findings, or expression are presented without adequate attribution — a definitional question that does not change based on how heavily the wording has been altered. Three distinct situations are worth separating clearly, because institutional response should differ across them:
- Properly cited paraphrase. Restating another author’s point in different words, with a citation to that author, is standard scholarly practice and is not plagiarism — whether the paraphrasing was done by hand or with an AI tool’s help.
- Uncited close paraphrase. Restating another author’s point in different words without a citation is plagiarism regardless of the similarity score it produces, and regardless of whether it was AI-assisted. A passage that scores 3% similarity but reproduces a source’s argument and structure without attribution has not become original by virtue of scoring low — see CASRAI’s Originality Report and Acceptable Similarity Percentage entries on why a low score is read as one input into a judgment call, not a clean bill of health on its own.
- AI paraphrasing used specifically to evade detection. Running someone else’s text (or one’s own previously published text, raising a separate self-plagiarism question) through a paraphrasing tool for the purpose of lowering a similarity score, while presenting the result as original work, compounds an underlying attribution problem with an intent to evade the institution’s own integrity check. Most institutional academic-integrity policies treat evasion intent as an aggravating factor, not a mitigating one.
Legitimate AI-assisted language editing vs. disguising unoriginal work
It matters that this is a genuinely different situation from AI-assisted language editing, which is widely accepted and, for many researchers, addresses a real and well-documented burden. CASRAI’s guide to AI proofreading tools for non-native English researchers covers that use case: a researcher uses an AI tool to polish the phrasing of their own original analysis and argument, generally with disclosure where a journal requires it, and generally alongside dictionary-term coverage such as AI-assisted writing.
The distinction that matters for an academic-integrity office is not “was AI involved” — it is whose ideas and words are being reworked, and why:
- Whose content. Polishing your own original sentences is language editing. Running someone else’s published text through a paraphraser and presenting the output as your own analysis is not language editing regardless of how much the wording changed.
- Purpose. Tools used to improve clarity and grammar in a researcher’s own writing serve a different function than tools used specifically to reduce a similarity-detection score on text whose substance originated elsewhere.
- Disclosure. Legitimate AI-assisted editing is increasingly expected to be disclosed. Undisclosed AI paraphrasing used to obscure a source is, by construction, not disclosed — disclosure and evasion are close to mutually exclusive in practice.
Undisclosed AI use as a separate integrity issue
Even where a paraphrased passage would not meet a traditional plagiarism definition — for instance, AI-generated text that does not derive from any single identifiable source — undisclosed AI use can still be an integrity problem in its own right, independent of similarity scoring. The Committee on Publication Ethics (COPE) position statement “Authorship and AI Tools,” published February 2023, states that AI tools cannot be listed as authors and requires authors who use generative AI in producing manuscript text, images, or in analyzing data to disclose which tool was used and how, typically in the Materials and Methods section or equivalent. COPE, together with the World Conferences on Research Integrity Foundation, the International Science Council, STM, and the Global Young Academy, is developing a cross-organizational “Global Reporting Standard for AI Disclosure in Research” through 2026, aimed at harmonizing what should be disclosed and in what format across journals and institutions.
For an academic-integrity office, this means the relevant question set is broader than “does this match a source.” It also includes: was AI used at all, was that use disclosed, and does the disclosure (or its absence) itself violate institutional or journal policy.
How detection has adapted: combining AI-detection with similarity-detection
Because similarity checking alone can miss AI-paraphrased text, detection vendors have added AI-writing-detection layers designed specifically to catch it. Turnitin launched an AI paraphrasing detection feature in July 2024 that is built to identify text that was likely AI-generated and then further modified using an AI paraphrasing tool or “text spinner” — examining writing patterns, style consistency, and linguistic markers that a resubmission-and-reword pass tends to leave behind, in addition to the direct-overlap matching a conventional similarity report performs. This runs alongside, not instead of, similarity checking: CASRAI’s guides to detecting AI-generated text and AI-detection accuracy cover what these AI-writing-detection scores can and cannot support as evidence.
That accuracy guide’s central caution applies directly here: an AI-detection or AI-paraphrasing flag, like a similarity score, is one input into an investigation, not a standalone finding. False positives are a documented, real risk — see CASRAI’s explainer on why a paper might say “AI detected” for the false-positive side of this same tooling. A responsible institutional process treats both a similarity report and an AI/paraphrasing-detection flag as evidence to be reviewed by a person with the full context of the submission, not as an automated pass/fail gate.
Institutional policy responses
A recurring pattern in how institutions have updated academic-integrity policy language to cover this gap: rather than only prohibiting “submitting someone else’s work as your own” (traditional plagiarism) or “submitting AI-generated work as your own” (undisclosed AI authorship) as separate categories, updated policies increasingly name the combination explicitly — using any tool, AI-based or otherwise, for the purpose of evading a required originality or authenticity check. That framing closes the gap a purely wording-based plagiarism definition can leave open, because it targets the evasion intent directly rather than relying solely on how much the final wording resembles the source.
On the operational side, the practical response most integrity offices have converged on is procedural rather than purely technical: run both a similarity check and an AI-writing/paraphrasing-detection pass, treat elevated results on either as a trigger for human review rather than an automatic finding, and evaluate the full submission — drafts, revision history where available, and a conversation with the student or author — before reaching a conclusion. This mirrors the general evidentiary posture CASRAI’s AI Detection Accuracy guide sets out for AI-writing detection specifically, applied to the paraphrasing-evasion case.
Frequently asked questions
Can Turnitin detect text that has been run through an AI paraphrasing tool?
Turnitin’s AI paraphrasing detection feature, launched in July 2024, is built specifically to flag text that appears to have been AI-generated and then further reworded with an AI paraphrasing tool, by examining writing-pattern and style signals beyond simple text-overlap matching. As with any AI-detection output, treat a flag as an input to review rather than a conclusive finding — see CASRAI’s AI Detection Accuracy guide.
If I use an AI tool to paraphrase a source and I still cite it, is that plagiarism?
No — properly cited paraphrase is standard scholarly practice regardless of whether a human or an AI tool did the rewording. The integrity problem arises when paraphrase (AI-assisted or not) is presented without attribution to the original source, or when AI paraphrasing is used specifically to lower a similarity score while concealing that the underlying content originated elsewhere.
Is using an AI paraphrasing tool the same as using an AI writing tool like ChatGPT?
They raise overlapping but distinct questions. Generating original-seeming text from a prompt raises authorship and disclosure questions covered in CASRAI’s Is ChatGPT (AI-Generated Text) Considered Plagiarism? guide. Paraphrasing existing text specifically to alter its wording raises the additional, more specific question of whether that rewording was used to evade an originality check on content whose substance came from elsewhere.
Does a low similarity score mean a paper is free of plagiarism concerns?
No. A similarity score measures textual overlap against a reference corpus; it does not measure originality of ideas or adequacy of attribution. See CASRAI’s Acceptable Similarity Percentage and Originality Report entries for why institutions treat a similarity report as one input reviewed by a person, not a pass/fail threshold.
