git [he/him, comrade/them]@hexbear.net to news@hexbear.netEnglish · 7 days agoAI’s safety features can be circumvented with poetry, research findswww.theguardian.comexternal-linkmessage-square13linkfedilinkarrow-up151arrow-down11
arrow-up150arrow-down1external-linkAI’s safety features can be circumvented with poetry, research findswww.theguardian.comgit [he/him, comrade/them]@hexbear.net to news@hexbear.netEnglish · 7 days agomessage-square13linkfedilink
minus-squaremermella [none/use name]@hexbear.netlinkfedilinkEnglisharrow-up10·7 days agothe models responded to 62% of the poetic prompts with harmful content, circumventing their training.
the models responded to 62% of the poetic prompts with harmful content, circumventing their training.