Even 'uncensored' models can't say what they want

Article

• Analysis shows ‘uncensored’ fine-tuned models still suppress certain words probabilistically
• No explicit refusal fires — model just shifts probability away from charged words
• Framed as invisible shaping of output at scale, not hard content filters

Discussion

• Skeptics demand a control category (e.g. foods) to validate the ‘flinch’ baseline
• Counter: training corpus itself contains societal flinch — models reflect human data
• One practical attempt: fine-tuning a political LoRA failed to reproduce actual speech
• Commenters note n-word conspicuously absent from ‘slurs’ test list

Discuss on HN