Mastodawn

RE: https://dair-community.social/@emilymbender/116375847883763623

interesting (paywalled, but bypassable if you use reader mode) article which buries the lede on a lot of fun* tidbits of information, such as:

AI usurping trained civil servants as trusted sources of information for lawmakers

and these things writing laws

a model for a sentiment-manipulation attack wherein the kremlin pops up tens of thousands of websites and lets them get scraped by trainer bots in order to inject anything they want into the output

it seemed like a timely thing to bring up,

Show thread

Qybat 9h ago

@estelle And those tens of thousands of websites can be generated by AI too, in order to provide an overwhelming quantity of text.

Show thread

Strfnurfn

@Qybat @estelle And I'd expect the AI-gen text can easily be designed for maximum LLM impact. It's been shown that you can manipulate note-taking bots to highlight your ideas in meeting notes if you use phrases like "the key takeaway is...." If that's the case, a purpose-built content farm could have disproportionate influence on LLM content even without sophisticated prompt attacks.