Post #3147159

2026-05-21 20:17 UTC

@colinmford@typo.social My bet is the highest bidders will be the ones with the lowest model costs. IE they would not have a model at all, just a string they want to inject into your output when they see some keywords in the text input.

Replies (2)

@adrake@sfba.social 2026-05-21 21:56

@theeclecticdyslexic@mstdn.social @colinmford@typo.social It only takes one rogue token to mess up a reply, so there's some interesting potential for adversarial strategies where you wait for some other bidder to start emitting their ad (having already sunk some money into it) and then you outbid them on any random token and emit gibberish (or ideally something damaging to their brand). Or you could run that against Google directly and buy cheap tokens in the middle of answers that aren't attractive to other bidders, again emitting gibberish. Bonus, once the answer is messed up, it's probably cheap to keep emitting more gibberish, making the failure even more apparent to the user.

Open ##3147160
@MisuseCase@twit.social 2026-05-21 23:24

@theeclecticdyslexic@mstdn.social @colinmford@typo.social You’re probably right and given how Google does not do any kind of due diligence regarding what paying customers put on their ad platforms - it could be malware! - they aren’t going to do anything to stop what you’re describing either

Open ##3147163