RE: https://infosec.exchange/@isotopp/116606041466845720

I thought this was a joke but it legitimately works.

@waldoj Does it? Or are the "couple of sentences of real text present on the corresponding URL" really just slop that looks plausibly like something you'd be likely to find there?

@dalias 🤷‍♂️ IDK, I haven't run an audit, but I'm running search terms where I'm well familiar with the topic and main websites, and the results were legit-looking.

I'd be interested to know what access that Gemini has to Google's index. It sure *should* have full access, but all sorts of stuff should be true that is not.

@waldoj @dalias The problem is that "legit-looking" is literally the job of a language model. :-)

I wouldn't claim it legitimately works unless you've actually spot-checked a few results, at the very least.

@timmc @dalias I mean, I'm not conducting a study here. I opened some results, they have the words.🤷‍♂️
@waldoj @dalias Ah, OK! That sounds like the spot-checking that I was referring to.