The question is framed wrongly. Plagiarism means passing off someone else's work as your own. A language model is not a person, and the text it produces does not belong to anyone in particular. By the definition, ChatGPT output is rarely plagiarism.
That does not make it allowed. Most institutions have written a separate rule about AI tools, precisely because the old plagiarism definition does not cover it. That rule is the one to read — not the plagiarism section.
Three things that can actually go wrong
1. Your institution's rules
They vary enormously: outright bans, "allowed for ideas, not for wording", "allowed if disclosed". Some places distinguish between examined work and ordinary assignments. Look it up in your course handbook rather than asking a classmate — the answer differs between departments at the same university.
2. The model sometimes reproduces text verbatim
Language models are trained on text from the internet, and on material they have seen many times they can reproduce phrasing nearly word for word. This happens most with definitions, legal text, well-known quotations and passages from heavily cited sources.
When it does, that is plagiarism in the classic sense — and you are the one who submitted it. Not knowing is not a defence that tends to work.
3. Invented sources
The most dangerous failure. Ask a model for references and you will often get sources that look entirely right: plausible author, plausible journal, plausible year, sometimes even a DOI. They simply do not exist.
A marker who looks up one source and cannot find it will check the rest. And a reference list with fabricated entries is a far more serious matter than a clumsy sentence.
Check every single source a model gives you. If you cannot find it in a database or on the publisher's site, it does not exist.
About AI detectors
Tools that claim to determine whether a text was written by AI produce a probability estimate, not proof. They are wrong in both directions, and studies have shown the errors are not evenly distributed: text by people writing in a second language is flagged as AI more often, because a simpler vocabulary and more regular sentence structure resemble model output.
Two consequences follow. If you are wrongly accused, a detector score is not evidence, and you can ask for it to be supported by something else. And conversely: a low score does not clear you if the text contains material copied word for word.
How to use it defensibly
- Read the rule for your subject before you start, not afterwards.
- Use it for the work around the text — structuring, explaining a concept you do not follow, finding counter-arguments, fixing language — rather than for producing the prose.
- Verify every source in a real database.
- Disclose it if your institution requires it. One sentence in the method section costs nothing and removes the entire argument.
- Run the text through a plagiarism check before submitting — that is the only way to catch a verbatim reproduction.
The simple test
Can you explain every paragraph of your work, without looking, to someone who asks? Can you account for why that particular source supports that particular claim?
If you can, the text is yours, whatever tools were involved along the way. If you cannot, the problem is not that you used a model — it is that you did not do the work.