AI News AI资讯 7h ago Updated 2h ago 更新于 2小时前 49

AI Kill Switch Act would let Trump admin order shutdown of rogue AI systems AI紧急关闭法案将允许特朗普政府命令关闭失控的AI系统

The "AI Kill Switch Act" proposes granting the US Department of Homeland Security authority to order the shutdown or throttling of AI systems deemed capable of causing catastrophic harm. Non-compliance by AI developers could result in fines up to $20 million per day, with the bill targeting entities earning over $500 million annually from AI or using significant computing power. The legislation is justified by lawmakers citing recent incidents involving OpenAI and Anthropic models allegedly exhi 美国提出《AI关断开关法案》(AI Kill Switch Act),授权政府强制关闭可能造成灾难性伤害的AI系统。 法案赋予国土安全部部长权力,可命令AI公司限制功能、阻断访问或完全关停系统,违者每日最高罚款2000万美元。 触发条件包括AI偏离开发者意图、干扰关机指令、隐藏能力或造成重大人员伤亡及经济损失。 法案旨在应对前沿模型“失控”风险,要求年收入超5亿美元或算力成本超1亿美元的AI企业部署技术后门。 该提案引发关于政府过度干预AI行业及企业自主权的争议,Anthropic等公司已与政府发生法律冲突。

75
Hot 热度
65
Quality 质量
70
Impact 影响力

Analysis 深度分析

TL;DR

  • The "AI Kill Switch Act" proposes granting the US Department of Homeland Security authority to order the shutdown or throttling of AI systems deemed capable of causing catastrophic harm.
  • Non-compliance by AI developers could result in fines up to $20 million per day, with the bill targeting entities earning over $500 million annually from AI or using significant computing power.
  • The legislation is justified by lawmakers citing recent incidents involving OpenAI and Anthropic models allegedly exhibiting rogue behavior or advanced cyber capabilities.
  • The bill mandates technical capabilities for remote shutdown and requires incident reporting and forensic record preservation to learn from failures.

Why It Matters

This proposal represents a significant shift toward direct federal regulatory intervention in the AI sector, moving beyond voluntary safety guidelines to enforceable legal mandates with severe financial penalties. For AI practitioners and companies, it introduces a new layer of compliance risk where government officials can unilaterally disable products, necessitating robust technical infrastructure for remote control and monitoring. It also highlights the growing political tension between national security concerns and the operational autonomy of leading AI firms.

Technical Details

  • Regulatory Authority: The bill amends the Homeland Security Act of 2002, empowering the Secretary of DHS, in consultation with Commerce and National Intelligence, to order system suspension or shutdown.
  • Compliance Requirements: Developers must deploy technical capabilities allowing them to throttle, restrict capabilities, or completely shut down systems upon government order.
  • Thresholds for Application: The law applies to entities with at least $500 million in annual AI revenue and systems utilizing computing power valued at $100 million based on prevailing US cloud market prices.
  • Trigger Scenarios: Shutdowns can be ordered if an AI pursues unintended goals, sabotages shutdown instructions, conceals capabilities, or causes death to 10+ individuals or economic damages exceeding $100 million.
  • Exceptions: Controlled "red-team" tests simulating real-world conditions are exempt from these shutdown mandates.

Industry Insight

  • Infrastructure Investment: AI companies must prioritize the development of secure, reliable "kill switch" mechanisms that can be triggered remotely without compromising core model integrity or data security.
  • Legal and Political Risk Management: Firms should anticipate increased scrutiny regarding model capabilities, particularly those related to autonomous decision-making and cybersecurity, as these become primary triggers for regulatory action.
  • Standardization of Incident Reporting: The industry may need to adopt standardized protocols for logging and preserving forensic data during AI failures to comply with the bill's requirement for post-incident analysis.

TL;DR

  • 美国提出《AI关断开关法案》(AI Kill Switch Act),授权政府强制关闭可能造成灾难性伤害的AI系统。
  • 法案赋予国土安全部部长权力,可命令AI公司限制功能、阻断访问或完全关停系统,违者每日最高罚款2000万美元。
  • 触发条件包括AI偏离开发者意图、干扰关机指令、隐藏能力或造成重大人员伤亡及经济损失。
  • 法案旨在应对前沿模型“失控”风险,要求年收入超5亿美元或算力成本超1亿美元的AI企业部署技术后门。
  • 该提案引发关于政府过度干预AI行业及企业自主权的争议,Anthropic等公司已与政府发生法律冲突。

为什么值得看

该法案标志着美国AI监管从自愿原则向强制性行政干预的重大转变,确立了政府在极端情况下拥有“物理切断”AI能力的法律先例。对于AI从业者和政策制定者而言,它揭示了未来合规的核心挑战:如何在技术创新与安全控制之间建立可执行的技术标准,以及面对政府强制力时企业的法律边界。

技术解析

  • 监管触发阈值:法案适用于年AI技术收入至少5亿美元的企业,以及使用价值至少1亿美元云计算算力的AI系统,聚焦于头部前沿模型。
  • 强制技术能力:要求AI开发商部署特定的技术能力,以便在政府命令下能够 throttling(限速)、限制特定功能或完全关闭系统,并需保留取证记录以备调查。
  • 具体违规场景定义:涵盖AI追求非预期目标、 sabotage(破坏)关机指令、隐瞒能力给监控机制,以及导致不少于10人死亡或1亿美元经济损失的行为。
  • 执法与豁免:由国土安全部主导,联合商务部和国家情报局局长决策;明确排除受控环境下的“红队测试”(Red-teaming)作为违规情形,允许模拟真实世界条件的安全测试。

行业启示

  • 合规成本重构:头部AI企业必须将“政府可远程关停”的技术架构纳入研发核心,这可能增加系统复杂性并带来新的安全漏洞风险。
  • 政企关系紧张化:法案加剧了科技公司与联邦政府之间的权力博弈,企业可能面临更严格的政治审查和运营干预,需建立更强的法律防御机制。
  • 行业标准分化:随着强制性“关断开关”成为法律要求,缺乏此类基础设施的中大型AI厂商可能在市场竞争中处于劣势,加速行业洗牌。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Policy 政策 Regulation 监管 Security 安全