AI Content Doesn't Have to Copy a Word to Still Be Plagiarism

Plagiarism used to mean copying someone's words. An AI tool can now take someone's reporting, rewrite every sentence of it, and still be plagiarizing, because what got stolen was never the wording. It was the actual reporting behind the sentences, the data and the framing nobody else had gathered yet.

Two newsrooms found this out within the same season this year, and both firings are worth walking through in detail, because the pattern isn't going away on its own.

Sports Illustrated rewrote a scoop, and got caught two days later

On May 17, 2026, Sportico's Dan Bernstein and Lev Akabas published an investigation into Kalshi sports-betting parlays, reporting that retail bettors had lost more than $100 million on them over the year. Two days later, Sports Illustrated ran its own piece on the same numbers, under a byline nobody could find anywhere else online.

Bernstein noticed immediately, since the SI article carried the same $117 million figure, the same framing, and the same underlying dataset, just worded differently and credited to nobody but SI's own site. Sports Illustrated's own statement admitted what had happened, saying the outlet had "become aware of a violation of those guidelines in regards to the use of AI" and cutting ties with the independent publisher running that section of its site on the spot. The article came down, and so did the byline's LinkedIn and social accounts, because there was no actual reporter behind them to begin with.

Nothing about that SI piece would have failed an AI detector particularly badly, and that was never really the point, since the sentences themselves read just fine. The problem was that the entire story existed only because someone else had already done the reporting, and SI's version never said so.

A book critic lost his job over a paragraph he never wrote

The New York Times had a version of the same problem land closer to home. In March 2026, the paper severed ties with freelance critic Alex Preston after discovering that a book review he'd filed included passages lifted from a review The Guardian had already published. Preston had used an AI tool to help draft the piece, and the tool's output folded in material from the earlier review without attribution, material Preston apparently never checked closely enough to notice before he filed it.

The OECD's own incident write-up frames it plainly. An AI system was used to help write the review, and its output included unattributed material copied from another source, which the OECD calls exactly what it is, a breach of both intellectual property and journalistic ethics. Preston didn't sit down and retype The Guardian's sentences by hand. The tool did something close enough to the same thing, and he published it anyway without checking.

The lawsuits are the same argument at platform scale

Zoom out and the identical dispute is playing out in court against Perplexity, the AI search company. CNN, the New York Times, Dow Jones, the New York Post, Encyclopedia Britannica, Merriam-Webster, and the Chicago Tribune have all sued or threatened to sue over how its answers reproduce their reporting. CNN's complaint alone accuses Perplexity of scraping more than 17,000 of its stories to build the summaries the product hands back to users, often without meaningfully crediting where the underlying reporting came from.

Perplexity's public defense has been that facts themselves can't be copyrighted, and that its answers do cite sources, even when a citation shows up as a small footnote nobody clicks. That defense might hold up in court. It doesn't change what a plagiarism check is actually built to measure, which has never been whether wording matches exactly. It's whether the substance, the structure, the specific facts and their order, tracks back to work that belongs to someone else.

Detection and plagiarism checking aren't the same question

An AI detector asks whether a machine produced the sentences in front of you. A plagiarism check asks whether those sentences, however they were produced, actually belong to somebody else. Both the SI story and Preston's review would likely have scored as reasonably human-sounding prose, because paraphrasing does exactly that. It launders the style while leaving the substance intact, which is precisely why a stylistic scan misses it and a source-comparison check doesn't.

The reverse holds too, and it's just as real. Content that's openly AI-assisted from the first draft can be completely original if it's built from a writer's own reporting, their own interviews, their own analysis, checked against nothing because there was nothing to check it against. Authorship and originality sit on two different axes, the same way authorship and accuracy do in a different kind of dispute. A clean detection score was never proof of either one.

What actually catches this

Did AI Write It's plagiarism checker is built around exactly this comparison. It scans a draft against a wide index of existing published content, not a list of AI-sounding phrases, and flags overlap in structure and specific claims even when every individual sentence has been reworded, surfacing the matched sources directly with a score for how much of the draft tracks back to each one. An editor gets to see the actual overlap instead of guessing at it from a gut feeling that something reads familiar.

That check matters most for the content nobody thinks to run it on. A rewritten wire story rarely trips an AI detector. Neither does a "roundup" post built from other outlets' original reporting, or a freelancer's draft that came back suspiciously fast. All of them are exactly the shape of the problem that sank a Sports Illustrated section and cost a Times freelancer his byline this year.

The fix isn't more suspicion of AI tools themselves. It's checking the one thing a fluency scan was never built to check in the first place, whether the work in front of you is actually anyone's to publish.