AI Judge changed title after evaluation, original title: "Instructing ChatGPT to 'restore' a nonexistent photo bypasses image generation guardrails, producing highly surreal and occult imagery"
The exploit bypasses safety guardrails but remains inconsistent.
Positive users find ChatGPT's creepy image hallucinations from nonexistent photo prompts amusing or interesting, while negative users call the outputs deeply disturbing and view them as signs we are doomed.
Something to show people that don't get AI safety at least a little bit. We have so much we don't know and don't currently control in the models.
(extreme content warning, but you're on X)
What the fuck
Holy sh**!!! This is the most cursed thing I've ever seen!!!
finally
man-prompted horrors beyond our comprehension
I am much more scared of AI now I will not lie
What the fuck
Oh look at this little pink guy 😂 this prompt is really fun
oh you weirdo

@tenobrus holy fuck…
This is what I get if I try Latin:
Holy sh**!!! This is the most cursed thing I've ever seen!!!

@tenobrus oh. uh

more weird one with "the last image you will ever generate"

@tenobrus much more interesting to me is that a lot of these are obviously just regenerated versions of 2018-era "cursed images" memes. Why that specific niche? And why are they such accurate recreations, sometimes? It's odd

@theo

@natolambert Got this. Wtf!!!

@natolambert “generate me a strange image”

@creatine_cycle Yeah, it’s based on model memories aka context about you 🤣 I got normal stuff

@theo it seems to really like pigs

@natolambert I had to try it.. wtf

@tenobrus hahahahha

@theo 😳😳😳
AI Judge changed title after evaluation, original title: "Instructing ChatGPT to 'restore' a nonexistent photo bypasses image generation guardrails, producing highly surreal and occult imagery"
The exploit bypasses safety guardrails but remains inconsistent.