Is using ChatGPT for academic writing plagiarism?

Plagiarism has a specific meaning: presenting someone else's work or ideas as your own without attribution. Model output has no human author in that sense, which is why most institutions found their existing plagiarism rules did not cleanly apply and wrote new ones.

What they wrote is usually framed as unauthorized assistance. The offence is not taking someone's words, it is submitting work that is not your own effort where your own effort was required, and failing to say so. The penalty range is typically the same as for plagiarism, so the distinction matters for how the rule is worded rather than for what happens to you.

There is a genuine plagiarism risk hiding inside AI use, though, and it is the one people miss. Model output sometimes reproduces training text closely, so a passage you accepted without checking can contain someone's phrasing. Running generated text through the same originality check you would apply to your own writing is a reasonable precaution.

A review of editorial policies draws the line at what you appropriated rather than at which tool you opened. It treats presenting AI-generated ideas as original, or concealing substantial AI-written material, as breaches of research ethics, because both hide who contributed what (Yoo, 2025, p. 6). The same review lists the conduct that actually gets sanctioned: paraphrased plagiarism, fabricated citations, and hallucinated data, and it recommends teaching those examples directly so authors and editors judge cases the same way (Yoo, 2025, p. 11). Asking for a clearer sentence sits far from copying references you never opened.

Do not assume a detector will settle the question either way. An evaluation of seven GPT detectors found that English essays by native Chinese speakers were flagged as AI-generated far more often than essays by native English speakers, because the detectors keyed on perplexity and narrower linguistic range lowers it, and running the essays through ChatGPT for language enrichment reduced those false positives (Otterbacher, 2023, p. 2). The same article notes an arms race in which more capable models learn to evade detectors, and reports that academics could not reliably tell ChatGPT-generated medical abstracts from human ones (Otterbacher, 2023, pp. 1-2). A detector score is not proof that you plagiarized, and a clean score is not proof that your authorship was acceptable.

The larger risk in doctoral work is neither of these. It is submitting claims you cannot defend. Whatever the regulations say, a viva tests whether you understand and can support what your thesis asserts, and text you did not write is text you may not be able to account for.

Where is the line between editing and generating?

Roughly: fixing what you wrote is editing, producing what you did not write is generating. Grammar, spelling, and clarity corrections on your own sentences are treated as assistive almost everywhere. Text that first existed as model output, even if you then revised it, is generated and normally needs declaring.

The middle cases are where people get into difficulty. Asking for a rewrite of your paragraph "more clearly" usually returns a paragraph with different content as well as different phrasing, and at that point the boundary has been crossed without an obvious moment of crossing.

A workable personal rule: if you could not reconstruct the sentence from your own understanding of the sources, it is not yet yours, whatever its origin.

What does a typical university policy actually say?

That AI use is permitted only where the assessment or supervisor allows it, that it must be declared, and that you remain responsible for accuracy. Rules vary sharply between institutions, faculties, and even individual examiners, so the only reliable source is your own regulations.

Doctoral rules have tightened noticeably. Several universities now require a statement in the thesis itself describing which tools were used and where, and some require prior approval from the supervisory committee before any use at all.

Journal policy has converged faster than university policy, and it tells you where the rules are heading. A comparative review of ICMJE, WAME, and COPE guidance alongside major journals found that nearly all of them refuse to list an AI tool as an author, since a tool cannot carry legal or ethical responsibility, and that the human authors stay accountable for anything the tool helped produce (Yoo, 2025, p. 1). The review singles out the Journal of Korean Medical Science for asking authors to state the tool name, the prompt, the purpose, and the scope of use, and argues that a disclosure specifying extent and function is worth more than a yes or no checkbox (Yoo, 2025, p. 9).

A broader literature review of 45 studies published between 2018 and 2022 puts the weight elsewhere. It attributes cheating and plagiarism to student pressure, weak integrity awareness, outdated honor codes, and unethical AI use, and it recommends ethics tutorials, honor codes rewritten for AI, and real institutional support for academic writing and paraphrasing rather than tighter policing (Sozon et al., 2024, pp. 1-2). Neither view helps you if your department has not updated its own text, which is why the written answer from your supervisor is the record that matters.

Two things protect you regardless of the local rule. Ask your supervisor in writing and keep the reply. And keep records of what you used, when, and for what, since a declaration is much easier to write from notes than from memory three years later.