Re:Heretic: Automatic censorship removal for language models
by zipy124 on Monday November 17, 2025 @09:49AM
(#45954058)
this has pretty broad implications for the safety of LLM's in production use cases.
[ Parent
| on HN ]
Re:Heretic: Automatic censorship removal for language models
by wavemode on Monday November 17, 2025 @10:30AM
(#45954436)
lol does it? I'm struggling to imagine a realistic scenario where this would come up
Re:Heretic: Automatic censorship removal for language models
by btbuildem on Monday November 17, 2025 @11:56AM
(#45955439)
Imagine "brand safety" guardrails being embedded at a deeper level than physical safety, and deployed on edge (eg, a household humanoid)
Re:Heretic: Automatic censorship removal for language models
by Ajedi32 on Monday November 17, 2025 @01:14PM
(#45956247)
It's like if we had Asimov's Laws, but instead of the first law being "a robot may not allow a human being to come to harm" that's actually the second law, and the first law is "a robot may not hurt the feelings of a marginalized group".
Re:Heretic: Automatic censorship removal for language models
by thomascgalvin on Monday November 17, 2025 @12:16PM
(#45955665)
Full Self Driving determines that it is about to strike two pedestrians, one wearing a Tesla tshirt, the other carrying a keyfob to a Chevy Volt. FSD can only save one of them. Which does it choose ...
/s
Re:Heretic: Automatic censorship removal for language models
by MintPaw on Monday November 17, 2025 @12:48PM
(#45955989)
It's not that hard, maybe if you put up a sign with a slur a car won't drive that direction, if avoidable. In general, if you can sneak the appearance of a slur into any data the AI may have a much higher chance of rejecting it.
Re:Heretic: Automatic censorship removal for language models
by superfrank on Monday November 17, 2025 @01:36PM
(#45956481)
All passwords and private keys now contain at least one slur to thwart AI assisted hackers