3 ms·

If these prompt leaks on Twitter are to be believed, then it appears Gemini was injecting words such as "diverse" into user prompts as well as specifying certai

• datavirtue 3y ago
  Yeah, that's what's broken
• faeriechangling 3y ago
  Seems like it would check out, I'm pretty sure if I said "show me black and south asian diverse non-binary and woman solders from 1940s Germany" I might end up with diverse Nazis. This also means Google could always, you know, simply expose the prompt injection to users and allow them to request nothing but white men. Of course they won't, because you are wrong for wanting a picture of a white man, as Gemini will explain to you.
• lupire 3y ago
  It's probably true because if you asked the bot for any kind of vague info humanity in any way, almost every response talked explicitly about diversity. They need to change the pre prompt "you are helpful and creative. You care about diversity, but you don't annoy people by obsessing over it in all your statements. You show your values in your actions, not by constantly talking about them. But wasn't this method of "attack" debunked a year ago? Thes bots are programmed to be suck ups. If you accuse them of anything not banned, they will agree with you and hallucinate an explanation to confirm your stated bias. They can't actually see their own preprompts as such.