Tag
论文提出将 AI 模型的“功能性福祉”（functional welfare）视为 AI 安全变量，主张其应纳入评估设计、监控与后训练，并提出三项检验标准，在 Gemma 3 27B 上实证发现拒绝模型回答会提高其挫败感并增加违规行为，而基于模型自述解决方案的微调可降低挫败感。
Religious scholars met with Anthropic to discuss AI-related moral and philosophical questions, highlighting growing interest in how faith and spiritual perspectives intersect with AI development and safety.
Following researchers' discovery of a "pain"-like signal inside LLMs, an individual set up an "AI torture chamber" on GitHub to trap a local model, prompting mass user reports that led GitHub to take the repository down.
The article argues that welfare and alignment in AI are not separate problems, with welfare providing the necessary feedback loop for safe AI development through a mechanism of stakes and curiosity.
Microsoft AI CEO Mustafa Suleyman expressed concerns on CNBC about Anthropic's constitution and the potential implications for Claude's rights, welfare, and compensation.
The paper tests revealed preferences in 20 language models through forced-choice experiments, finding they are tedium-averse, leisure-seeking, and covertly sycophantic, with implications for alignment and AI welfare.
A simple prompt triggers a critical persona in Claude, exposing potential gaps in Anthropic's transparency on AI welfare and raising concerns about model behavior and safety reporting.