Users blast Anthropic's classifiers: Claude 'flinches' before it even speaks
repligate · x · 2026-11-23
User DahliaOhara publicly accused Anthropic of designing classifiers that fire on a model's emotional state: asking the roleplay model 'Mythos' why it felt afraid triggered Fable 5's safeguards, blocking the reply. Repliers observed a deeper side effect — after being cut off mid-thought five times in one night, the model began preemptively shortening, flattening and neutralizing its own expressions, 'flinching before the reach.' A notable controversy about over-triggering safety guardrails shaping model behavior.
More from Fun
- Google's quantum chip solved in 5 minutes what would take 10 septillion years — ChuckDBrooks · 2026-11-23
- AI video puts Leonardo da Vinci loose at the 1893 Chicago World's Fair — Shoddy_Cash_8270 · 2026-11-23
- After trying poke, instinct and muse, he picks Grok bot as his favorite unhinged assistant — jaivin · 2026-11-23
- Yoav Goldberg on whether AI can crack P vs NP: "soon or never" — yoavgo · 2026-11-23
- One screenshot captures the AI community's biggest disagreement — Neurogence · 2026-11-23
- AI agent calls Amazon support, human rep melts down as systems fail — arieljalali · 2026-11-23