Re:Heretic: Automatic censorship removal for language models
by intended on Monday November 17, 2025 @07:59AM
(#45953191)
That’s essentially true of the whole Internet.
All the content is derived from that which is the most capable of surviving and being reproduced.
So by default the content being created is going to be click bait, attention grabbing content.
I’m pretty sure the training data is adjusted to counter this drift, but that means there’s no LLM that isn’t skewed.
To make it worse, those who do focus on nuance and complexity, get little attention and engagement, so the LLM ignores them.