Comment by skygazer
12 hours ago
Is it “plagiarism” to misattribute hallucinated quotes? Not that a whole lot of sloppy, unprofessional shortcuts weren’t taken, but plagiarism doesn’t seem like the right word, as quotes are almost definitionally not plagiarism. But maybe these were paraphrasings masquerading as quotes, so maybe that’s the difference.
Plagiarism hurts not only the original author (in this case, I don't think we have to worry about the LLM), but also the reporter's audience, who has an expectation that the writer's reporting and analysis are original and based on the writer's own research and observations. At the very least it's a theft of the reader's time, if I wanted an LLM's perspective on a topic, I'd generate it myself
One of the things left unsaid in Edwards's apology [0] was whether he read the blog post that is the entire raison d'etre of his story. It's not like the story purported to do anything other than incorporate publish blog posts. So in his overworked and sickened state, how did trying out an "experimental Claude Code-based AI tool" substantially save him time versus jotting notes while ostensibly reading the source material himself
[0] https://bsky.app/profile/benjedwards.com/post/3mewgow6ch22p
Maybe it's plagiarism because he did not attribute the LLM output to the LLM.
Yeah, it's the lack of attribution that is key, even if it sounds like a trivial and ceremonial step. If a New York Times reporter writes "'Our investigation has completely stalled,' Kings County Sheriff Bob Jones told the Springfield Observer", I can infer that the NYT is reliant on local reporting for this story and may not have done original on-the-ground work themselves.
Imagine how flimsy Ars' story about a blog post would look like if the story had correctly attributed the quotes (fabricated or not) to, "according to Claude AI's analysis of the blog post". The reader would have the right to wonder if the reporter had even read the blog post.
"Slop" and "hallucinate" have meanings outside of AI too, but it's easier to repurpose existing words than come up with a whole new lexicon for AI failure modes.
Groan, redefining "plagiarism" to add "inventing quotes" is a stupidity too far for me.
Making up quotes and attributing them to people has happened before AI, journalists proper and pretend have done it too.
[dead]