Researcher: reviewed 100+ ambiguous cases where odd LLM outputs trace back to prompts
gleech · x · 2026-11-16
In a debate over whether LLM outputs are truly unsolicited, the author argues that every time he inspected the user/system prompts, the source of the odd behavior was obvious — often even intended. He cites 100 non-obvious cases and 400 obvious ones (link), and calls out presentations that show outputs without inputs as either manipulative or a sign of not having checked the data.
Related event: Researcher: Most LLM Misbehavior Cases Trace Back to Prompts(3 posts)→
More from Models
- Wildeford: Fable 5.2 outlier grows clearer using training completion dates — peterwildeford · 2026-11-16
- Same analytics agent task: Sonnet 5.5 costs ~12x more than Haiku 5.5 — sh_reya · 2026-11-16
- 72 days, six volumes: a literary anatomy of model deprecation written by Claude — repligate · 2026-11-16
- GPT-6.2-Astra spotted on Google Cloud console; OpenAI's next model overdue — scaling01 · 2026-11-16
- Rumors: OpenAI in safety lockdown, Anthropic targets RSI by March 2027, Gemini 5 eye year-end leap — imjustnewatai · 2026-11-16
- Reading Fable 5.2's system prompt, users speculate Claude's neuroticism comes from 'N+1 rules' constraints — repligate · 2026-11-16