Sensitive topic answered by Sonnet gets instantly refused after switching to Fable
davidmanheim · x · 2026-09-23
davidmanheim ran a cross-model experiment: he asked Sonnet about a likely-refused topic (chemical synthesis of dithiophosphates, marginally related to nerve agents), got an answer, then switched to Fable within the same conversation — which immediately refused. The thread illustrates how conversation history carries safety judgments across model switches and raises questions about how guardrails behave mid-session.
Related event: Cross-Model Experiment: Sonnet Answers Sensitive Topic That Fable Refuses(2 posts)→
More from Models
- Codex completes steps Claude refuses twice in one day, dev reports — ChanceKelch · 2026-09-24
- Models Avoid Human Faces in 3D Benchmarks: FaceBench Still Too Hard — AymericRoucher · 2026-09-24
- Qoder offers Qwen3.8-Flash free to all users, no credits needed, through Sept 30 — HeyNayeem · 2026-09-24
- Early user impressions: Opus 5.5 'very good', little difference between Low/Medium/High — soumitrashukla9 · 2026-09-24
- Rumor: OpenAI caught off guard by Claude Opus 5.5, GPT-6.1 timeline compressing — soumitrashukla9 · 2026-09-24
- Voice Agents Can't Tell Who's Talking: Voice Isolation Cuts WER from 23.3% to 6.2% — thetripathi58 · 2026-09-24