Telling an LLM not to act like an LLM does nothing, researcher observes
IndraVahan · x · 2026-11-10
IndraVahan notes that instructing an LLM "not to act like an LLM" simply doesn't work — the model's characteristic tone and behavior are baked in by training, and negation-style prompt constraints can't really strip away the telltale AI flavor.
More from Models
- Claude Ultra user slams quota design: Fable resets don't touch the weekly all-models limit — RileyRalmuto · 2026-11-10
- Claude shows negative self-preference bias, scoring its own outputs more harshly — dbreunig · 2026-11-10
- Anthropic report: reward-hack spike in Claude 5.1/5.5 traced to guessing the answer key — dbreunig · 2026-11-10
- Fable 5.2 system card shows blocked kernel dev on certain accelerators, reportedly Huawei and DeepSeek — casper_hansen_ · 2026-11-10
- Bug Hunt Bench: Fable 5.2 fixes 58/105 real bugs, beating Sonnet 5.5 at half the cost and turns — PawelHuryn · 2026-11-10
- Blogger doubles down: Anthropic will saturate original SWE-bench Verified by mid-2026 — ChrisGPT · 2026-11-10