Post

Thomas Spooks 🎃👻💀
@thomasfuchs.at

LLMs would actually be reasonably good at figuring out if something is fiction or satire, especially if combined with some other data, like what sites link to the information and what other content is on the site. Of course that would be actually useful and isn’t driving “engagement”.

0 reposts9 likes

◻◊Happy
@stupidcomputer.bsky.social

It's just... of course there have always been fun problems with search engine algorithms and ways to game them. But seeing a google feature use a quote from an article about how that quote is not true feels like a new layer of dumb problems (at least it cites its source I guess).

Dilong
@dilong.bsky.social

Why do you say this? Satire can be hard for humans to parse and fictional docs are often intentionally written with identical verbiage and structure to non-fiction ones. You need context AI doesn't have. The site links idea is better but still doesn't prevent an April 1st spaghetti farm mistake

Thomas Spooks 🎃👻💀
@thomasfuchs.at

Because for judging texts or statements that are false on their face, generic LLMs have relatively high accuracy (definitely not perfect, though). That's because a lot of the text corpus they ingested will directly contradict claims in satirical or otherwise fictional writing.

Thomas Spooks 🎃👻💀
@thomasfuchs.at

Also true. Also, Google has a lot of meta-information they could tap into, that's more useful in case (like the reputation of sites linking to it, is it linked to from websites dealing with fiction or entertainment, was it published around April 1st, etc.).