logoalt Hacker News

Kotlopou • today at 1:53 AM • 1 reply • view on HN

Case in point: Pangram says this article is 80% AI-written. [1] I only started suspecting in the very last paragraph ("Michel Dufrénoy crosses Verne's frozen Paris carrying almost nothing: a memory of books, a love of poetry, a handful of dying flowers and the name of someone he loves." is a vaguely poetic-sounding structure that I've already seen dozens of times in the last year).

Ignore the debates about quality. This means that this article's message is a lie; all of the opinions expressed here come not from a human, perhaps an inhabitant of Paris, but from a machine. The salvation suggested in the form of Parisian 1960s poets does not come from someone who likes to read them, but from a machine reporting on its training data. (Here it's not controversial -- what else can an LLM know about a poet besides what has been written about them?)

I feel like exploring a cave now.

[1] The 20% flagged as human is mostly direct quotations from the book, which is a nice sign that Pangram isn't making things up randomly.


Replies

cubefox • today at 7:18 AM

Also note that HN seems to check comments using Pangram now, so citing LLM text (or article text that was written by an LLM) can get it (the comment) shadowbanned. In your case the citation is just a small fraction of the whole comment though.

➕ show 1 reply