Plagiarism, Paraphrase, and Where AI Output Sits

“Is AI writing plagiarism?” bundles three separate questions with three different answers: whether it breaks the rules of an institution, whether it infringes anyone’s copyright, and whether it breaks the promise a byline makes to a reader. Answered apart, each becomes tractable.

Three questions wearing one word

They come apart in every direction. Generated text submitted as coursework can be an academic offence while infringing nobody’s copyright. A verbatim paragraph from a paper can infringe copyright and not be plagiarism, if it was quoted and cited properly. A ghostwritten column breaches no rule of either kind, and is an editorial problem in some contexts and completely normal in others.

The academic question: rules, not law

Academic misconduct rules are about representation. The classic offence is presenting someone else’s words or ideas as your own without acknowledgement. The newer offence, which most institutions now handle separately, is submitting work you did not produce — whether it came from a contract cheating service or from a model.

Two consequences are consistently missed. First, this is a rules question rather than a truth question: whether the use was permitted is decided by the specific assessment brief and the institution’s regulations, and those vary enormously. Some modules require a declaration of tool use, some permit it for drafting only, some prohibit it outright. Read the brief, not just the general policy.

Second, an originality report is not the test. Text-matching software compares a submission against a corpus, so generated prose that matches nothing scores clean and can still be misconduct under the rules on unauthorised assistance. In the other direction, detectors that claim to identify machine authorship produce a statistical judgement about style rather than evidence about what happened, which is why how institutions should treat a detector score is contested everywhere it has been tried.

The copyright question, in two halves

People asking about copyright are usually asking one of two quite different things, and conflating them produces most of the confusion in the public argument.

Half one: does the output infringe someone else’s rights?

Copyright protects expression, not ideas, facts or styles. Writing about the same subject as an existing article is not infringement; reproducing a substantial part of its actual expression can be. That principle predates the technology by centuries and does not change because of how the text was produced. A copied paragraph is a copied paragraph regardless of what typed it.

Practical exposure is therefore concentrated in the cases where output reproduces protected expression closely — uncommon for ordinary prose, less uncommon for short canonical texts. It is not a defence that you did not know: liability for infringement generally does not require intent, though intent affects remedies. Whether training on copyrighted material was itself lawful is a separate dispute being litigated in several jurisdictions, and it is covered on the training data page.

Half two: do you own the output?

Here the answer depends heavily on where you are, and two published positions are worth knowing because they differ so plainly.

The United States Copyright Office published guidance in 2023 on registering works containing material generated by artificial intelligence, stating that copyright protects material of human authorship and that an applicant must identify and disclaim machine-generated content that is not the product of human authorship — with the protectable part being the human contribution, such as selection, arrangement and modification (copyright.gov/ai). The United Kingdom took a different route long before any of this arose: section 9(3) of the Copyright, Designs and Patents Act 1988 assigns authorship of a computer-generated work with no human author to the person who made the arrangements necessary for its creation (legislation.gov.uk).

The practical consequence is the same under either approach: the more of the final expression is yours — your selection, your arrangement, your rewriting — the stronger your position, and a piece assembled from unmodified generated paragraphs is the weakest case you can construct. If you licence work to clients this is not academic. A client who needs to own the deliverable outright is buying something you may not be able to convey, and that conversation belongs before the contract rather than after delivery.

The editorial question: what a byline promises

The editorial question is not about rules or rights. It is about whether the reader has been misled about the provenance of the work, and it is the one that decides whether people keep reading you.

It is also the only one of the three a competent writer feels rather than looks up. A byline on a first-person essay promises that a person had those experiences. A byline on a review promises somebody used the product. A byline on a reported piece promises the reporting happened. None of those promises is broken by a spellchecker and all of them are broken by generating the substance, which is why the disclosure threshold is drawn at contribution to substance rather than at tool use.

Paraphrase deserves its own note, because it is where the three questions diverge most. Rewriting a source in your own words without citation is a clear academic offence and a clear editorial one, and is frequently not copyright infringement at all, because the ideas were never protected. A model is unusually good at producing exactly that artefact: a fluent paraphrase with the citation stripped. This is why the workflow that grounds every claim in a quoted span is a plagiarism control as much as an accuracy one — a claim that arrives with its source attached cannot lose it silently.

Regurgitation: the one mechanical risk

There is a real mechanism underneath the copyright worry and it is narrower than the debate suggests. Models sometimes reproduce training data verbatim, and it happens most where a passage appeared many times in the corpus and where the prompt closely cues it: lyrics, famous opening lines, licence boilerplate, widely syndicated copy, popular code snippets.

• Duplication in the corpus. Text that appears in thousands of documents is far more likely to be reproduced than text that appears once.
• Short, canonical, heavily cued passages. Asking for “the opening lines of” something is close to asking for a lookup.
• Greedy decoding on a strongly cued prompt. Temperature zero follows the single most probable continuation, which is precisely the memorised one wherever memorisation exists.

The control is boring and effective. For anything you will publish, run the draft through ordinary text-matching against the web before it goes out. That catches genuine reproduction whatever its origin — a model, a careless researcher, or a copy-paste from a tab you forgot — and it is the check a publisher would run anyway.

What to actually do

1. Work out which of the three questions you are in. Coursework is the first. A client deliverable is usually the second. A byline is the third. Many situations are two at once; almost none is all three.
2. For the academic question, read the specific assessment brief, and if it is silent ask in writing before you submit. A written answer is the whole defence.
3. For the copyright question, keep the human contribution real and documented: your outline, your sources, your edits, your selection. Keep the drafts and their timestamps.
4. For the editorial question, apply the substance threshold, disclose above it, and name the person who verified the piece.
5. In all three, ground every claim in a source you opened yourself. The workflow that does this happens to solve most of the other two problems as a side effect.