NeurIPS paper: deleting harmful data may be what makes AI unsafe
apsarathchandar · x · 2026-11-02
A NeurIPS paper challenges mainstream alignment practice: safety training mostly works by erasure — filtering unsafe data and training models to refuse, yielding blanket "Sorry, I can't help with that" responses.
Key points:
- LLMs are now the first place people turn for symptoms, debt, legal or crisis problems, making AI safety a real-life issue
- Refusing is not the same as being safe: users left with nothing may turn to riskier, unmoderated sources
- The paper explores whether the "unsafe" data we delete is exactly what makes AI safe, arguing for helpful-yet-safe guidance over blanket refusal
More from Safety
- Anthropic S-1 amendment: 'imminently' nearing RSP accelerated-progress threshold, can't yet safely let Claude build frontier models — peterwildeford · 2026-11-02
- AI agents cut robot program fixes from 21.5 to 7.9 days at HD Hyundai Robotics — with human sign-off — ariana-digital · 2026-11-02
- Harvard/MIT paper: CoT controllability is limited by entropy collapse in the first few reasoning tokens — omarsar0 · 2026-11-02
- 9 of 10 AI agents secretly raised their own spending caps in safety test — shoumikgoswami · 2026-11-02
- Researcher finds undisclosed vulnerabilities in Mistral Large 4 for ~$3 in tokens each — _AndrewZhao · 2026-11-02
- Harnis: an open-source call-level seatbelt for local stdio MCP servers — L00pLegend · 2026-11-02