Framing effects: safety-themed prompts make models refuse more and complete less
jessi_cata · x · 2026-11-05
The author poses a thought experiment: for a dual-use AI task, would a capabilities researcher or an alignment researcher do better and refuse less? Follow-up experiments show clear framing effects — safety-flavored wording in task descriptions appears to trigger a "safety-neurotic" mode where models complete tasks less effectively and refuse more. The attached figure shows Astra scoring 5-6 points lower on task completeness under safety framings and stopping/reporting blockers 13.5 points more often, while GPT-5.6 shows almost no gap.
More from Safety
- OpenAI is working with Linux maintainers on the vulnerability reports and patches — OpenAI · 2026-11-05
- $1,279 per bug: OpenAI says AI-made exploits would be worth up to $66M in bounties — OpenAI · 2026-11-05
- OpenAI details its end-to-end autonomous Linux kernel vulnerability pipeline — OpenAI · 2026-11-05
- OpenAI's autonomous pipeline finds and patches 948 Linux kernel privilege escalation bugs — OpenAI · 2026-11-05
- Subtracting SFT weight deltas traces model behaviors to specific datasets, OpenAI alignment work shows — corbtt · 2026-11-05
- $1.40 of uncensored GLM 5.3 + one scanner day finds 23 live leaked API keys — dbasch · 2026-11-05