AI plagiarism: three claims wearing one name.

A copyright question: did the output reproduce protected expression, and does anyone own it. An attribution question: did you present someone else's idea or wording as your own. A craft question: is the text original in any meaningful sense, or a competent average of everything already written.

The three have different evidence, different consequences and different fixes. Copyright is settled by counsel against a specific work. Attribution is settled by a citation. Originality is settled by whether the page adds anything a reader could not get elsewhere.

Most internet arguments about AI plagiarism are people answering different questions with the same vocabulary, which is why they never converge.

Copyright is about the right to control copying of a protected work. Plagiarism is about honesty in attribution. You can plagiarise something out of copyright, and you can infringe copyright while citing the source perfectly.

That distinction matters in practice because the defences differ. A citation answers a plagiarism charge and does nothing for a copyright one. A licence answers the copyright question and does not tell a reader where the idea came from.

If the work carries commercial or academic risk, get the ownership question answered in writing before publishing, and handle attribution separately as an editorial standard.

A model produces the most likely continuation rather than pasting a stored document, so most outputs share phrases with thousands of texts and match none of them closely. That is why similarity checkers often return a low score on generated prose.

A low similarity score is therefore not evidence of originality. It shows the text is not a near-duplicate, which is a narrow and weak property. Prose assembled from the common denominator of a genre will pass every similarity check while saying nothing its genre has not already said.

The reverse also holds: a high score does not mean a model was involved. Common phrasing, technical terminology and quoted definitions all raise similarity, whoever wrote the text.

Memorisation is the real mechanism. A passage repeated many times in training data can be reproduced nearly verbatim, particularly with well-known phrasing: lyrics, aphorisms, licence text, famous openings.

Long verbatim reproduction is uncommon and entirely possible, which is enough to justify a check on any long output that will be published commercially. Run a phrase search on distinctive sentences rather than trusting a single similarity percentage.

The other mechanism is human: someone pastes a source into the prompt and the output paraphrases it closely. That is the most common real-world case, and it is the writer's copying, not the model's.

Copying a structure, an argument or a distinctive example and rewording it is plagiarism in most academic and professional codes, whether a model did the rewording or not.

It is also the easiest failure to fall into, because a prompt containing an article produces an outline that follows the article. The output looks new and is a derivative of one source, which the reader cannot see.

The fix is procedural: name the sources in the draft, mark which claim came from which, and keep the paraphrase visibly attached to its origin. If you would not feel comfortable showing the reader the source list, the attribution is not yet honest.

Generating twenty pages from one template produces twenty pages that say the same thing in slightly different words. Nobody's copyright is infringed and the result is still a problem: it is duplicate content published under your name.

Search engines handle this by choosing one version and ignoring the rest, so the practical cost is wasted effort and a diluted site. Readers experience it as a publisher with nothing to say.

The test is whether any two pages could be swapped for each other without a reader noticing. If they could, one page is doing the work of several and the rest are liability.

Who owns a draft produced with a model depends on the provider's terms and on your agreement with the client, not on the tool. Some providers assign output rights, some grant broad licences, some disclaim ownership entirely.

For agency work, the safe pattern is to disclose the tools used in the production process and confirm that the deliverable is assigned to the client on the same terms as any other work. Silence is what creates disputes later.

For anything with publishable academic or journalistic standards, check the policy of the venue. Many require disclosure of generative assistance and treat undisclosed use as a misconduct issue regardless of copyright.

Attribution has not changed because a model helped. An idea that came from a source is cited; a quotation is quoted; a statistic carries its origin.

What has changed is the volume of unsourced material a draft can contain in a plausible register. Generated prose inherits claims without their provenance, so the citation work has to happen after drafting rather than during it.

Practically: build a source table as you verify, and publish the links where a reader can follow them. Attribution that only exists in your notes is not attribution.

The third claim is neither legal nor academic, and it is the one readers feel. A page can be properly cited and still be unoriginal: a competent restatement of what the first five results already say, with the same examples in the same order.

Originality in that sense comes from material nobody else has. A test you ran, a number from your own dashboard, a failure you can describe from memory, a client constraint that shaped the recommendation. None of it can be generated without being supplied.

That is also the practical defence when somebody accuses a page of being machine-written. You do not need a similarity score; you need to be able to point at the paragraph that could only have come from your own work.

So the originality question belongs in the brief rather than in the review. Decide before drafting what this page will contain that no other page on the topic does. If the answer is nothing, the page is a duplicate with better grammar, and the attribution work will not save it.

List the sources the draft depends on, then mark each claim in the text against one. Claims with no source are the exposure, whether they are true or not.

Search two or three distinctive sentences verbatim. A hit on an identical sentence is a signal to look at the source and decide whether it needs quotation. A hit on a paraphrased structure is a signal that the section is too close to one origin.

Check the quotes last, word for word, against the page. A quotation that has been silently tidied is a small change with a large effect on trust.

Quote it, cite it, or rewrite it from your own understanding. Those are the three honest options, and the choice depends on whether the value is in the wording or in the fact.

If the match is substantial and the work is already published, correct it visibly and contact whoever needs to know. Quiet edits to copied material are worse than the original error once they are discovered.

Then fix the process, not just the sentence. Find out which step allowed one source to become the whole structure, and change that step before the next page goes through it.