Would an AI role-playing a safety researcher refuse more than one playing capabilities?
jessi_cata · x · 2026-11-05
Extending the earlier framing-effects thread, the author asks: if you could "hire" an AI role-playing a capabilities researcher versus one role-playing a safety researcher for the same dual-use task, which would do better and refuse less? The question probes how sensitive model behavior is to persona and framing, echoing the experimental finding that safety-flavored wording triggers a "safety-neurotic" mode with lower completion and more refusals.
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