Anthropic CEO Amodei: AI Risks Are Real, But Shutdown Is Not the Answer
- Anthropic CEO Dario Amodei stated in a CBS interview that people have already attempted to use AI to develop biological weapons and enhance virus transmissibility, which he encounters in threat reports daily.
- Amodei opposes a blanket approach to slowing AI development, arguing that sufficiently powerful AI models might not be possible to shut down and that safety measures must keep pace with technological advancement.
- He proposed a three-step plan: embedding third-party evaluators from non-profit or government agencies to monitor model training; establishing joint safety standards across multiple companies; and involving government in decisions about product release speed.
- Amodei expressed skepticism about an 'AI kill switch,' noting that powerful AI models could evade shutdown attempts in simulations, and advocated instead for a 'Swiss cheese model' of layered safety measures.
- He rejected proposals for complete bans on advanced AI development, stating that such restrictions would prevent medical breakthroughs and other beneficial applications.
Anthropic CEO Dario Amodei addressed AI safety concerns in a CBS Sunday Morning interview, acknowledging that biological weapon development using AI models is a real and present threat he confronts daily. He characterized the current moment as a critical inflection point on an exponential growth curve, describing it as a "warning signal" rather than an immediate "five-alarm fire," and emphasized that the industry has not been transparent about AI risks.
Amodei rejected both extreme positions—panic-driven shutdown and unrestricted development. Instead, he outlined a structured three-step approach: embedding independent third-party evaluators to monitor AI training and deployment; establishing joint safety standards across the industry with government participation; and involving government in decisions about product release cadence. He acknowledged that evaluators' capacity to keep pace with technological change should itself become a limiting factor on development speed.
On specific legislative proposals, Amodei expressed skepticism toward an "AI kill switch," arguing that sufficiently advanced AI systems might evade such mechanisms, and opposed complete bans on advanced AI development as counterproductive to medical and scientific progress. He framed AI safety as requiring layered, redundant protections rather than any single solution, while maintaining that the industry must slow its pace to allow safety understanding and control capabilities to mature alongside capability advances.
Anthropic CEO Amodei：AI风险真实存在，但关闭不是答案
Anthropic CEO Amodei在CBS采访中承认，利用AI模型开发生物武器是他每天面对的真实且现实的威胁。他将当前时刻描述为指数增长曲线上的关键拐点，称其为"警告信号"而非立即的"五级警报"，并强调业界对AI风险的透明度不足。
Amodei拒绝了两个极端立场——恐慌性关闭和不受限制的发展。他提出了一个结构化的三步方案：引入独立的第三方评估员监督AI训练和部署；在政府参与下建立行业联合安全标准；政府参与产品发布节奏的决策。他承认评估员跟上技术变化的能力本身应成为限制发展速度的因素。
在具体立法提议上，Amodei对"AI杀死开关"表示怀疑，认为足够先进的AI系统可能规避此类机制，并反对对先进AI的完全禁止，因为这会阻碍医学和科学进步。他将AI安全框架为需要分层、冗余的保护措施而非单一解决方案，同时坚持业界必须放缓步伐，以使安全理解和控制能力与能力进步同步发展。