
谷歌DeepMind掌门人提议美国主导设立AI标准机构,应对通用人工智能临近风险
哈萨比斯呼吁建立类似金融业监管局的独立实体,对前沿AI模型进行安全测试,此举正值美国政府加强对AI出口管制之际,引发全球治理路径之争。
谷歌DeepMind首席执行官德米斯·哈萨比斯本周公开发文,提议由美国牵头建立一个全新的AI标准机构,对“前沿级”模型进行国家安全风险评估,并设定公开发布前的安全门槛。这一倡议的直接背景是,特朗普政府近期以国家安全为由,先后限制OpenAI和Anthropic的模型对外提供访问,其中Anthropic的Claude Fable 5与Mythos 5因越狱漏洞被勒令暂停所有境外访问并下线逾两周。该事件在欧美和亚洲政策圈引发震动,被多位科技企业高管视为美国现行AI监管框架缺位的“警钟”。
哈萨比斯构想的机构以美国金融业监管局(FINRA)为蓝本——一个由国会授权、行业资助、独立运作的非营利监管组织。新机构将设立由顶尖技术专家和开源社区代表组成的董事会,负责制定并动态更新前沿模型的能力基准,任何达到“前沿级”门槛的模型,无论来自美国还是海外开发商,均需在发布前至少30天自愿提交审查;待测试体系成熟后,合规将转为强制。哈萨比斯强调,该机构须由AI产业提供主要资金,以吸引世界级人才并配备大规模测试所需的算力资源。
这一方案凸显了美国科技界在AI治理路径上的分歧。OpenAI首席执行官萨姆·奥尔特曼主张通过国际联盟和全球条约体系来协调监管,而哈萨比斯认为等待全球共识进程过慢且不切实际,美国应单方面建立框架,凭借其市场和技术地位迫使其他国家——包括中国——为保持进入美国市场而接受这套标准。亚洲政策观察人士指出,随着中国DeepSeek、Z.ai等开源模型快速缩小与美国的差距,美国主导的标准机构若落地,可能演变为事实上的全球技术准入门槛,对亚洲AI开发商形成新的合规压力。
哈萨比斯在文中同时描绘了一幅谨慎乐观的技术前景:通用人工智能(AGI)可能仅剩数年之遥,其冲击规模或相当于工业革命的十倍,并以十倍速度发生。他呼吁哲学家、经济学家等社会各界共同参与定义这一技术变革的价值观与目标,而非仅由技术专家决定。这一表述被欧洲评论界视为对当前AI治理讨论过于集中在安全与风险领域的一种补充,试图将更广泛的人类境遇议题纳入议程。
接下来值得关注的节点是,美国商务部及国会是否会就这一行业资助、联邦监督的混合模式展开实质性讨论。此前Anthropic首席执行官达里奥·阿莫迪亦曾呼吁设立联邦机构对强大模型进行部署前审查,但白宫尚未明确表态。哈萨比斯的提议能否从个人倡议转化为政策选项,取决于华盛顿在即将到来的选举周期中,能否在创新激励与国家安全之间找到可操作的平衡点。
| 伊朗及相关媒体 | −0.10 | neutral |
|---|---|---|
| 中国媒体 | +0.20 | neutral |
| 大西洋/英语圈媒体 | 0.00 | neutral |
| 印度及南亚媒体 | −0.30 | critical |
The Google DeepMind CEO issues a cry of alarm: time is short and the Trump administration has already begun restricting rival models. The proposal is a necessary response to an imminent threat.
It uses emotional language ('shocking warning', 'precious window') and links the proposal to Trump administration actions to create a sense of urgency and necessity.
It omits the technical details of the proposal (FINRA model) and international cooperation, focusing solely on the alarm and US actions.
China recognizes the unique position of the United States to lead AI regulation and supports the creation of an independent agency modeled on FINRA. The proposal is seen as a pragmatic step.
It adopts a technical and descriptive tone, explicitly citing the FINRA model and the US economic position to legitimize the proposal as rational and feasible.
It omits the CEO disagreements and geopolitical tensions, presenting the proposal as a technical solution without conflict.
The West reports Hassabis's warning with detachment, presenting facts without taking a stance. The proposal is described as one of many in the ongoing debate.
It uses standard journalistic register, citing sources and statements, balancing urgency with a measured tone and avoiding judgment.
It does not delve into CEO disagreements or geopolitical context, keeping the focus on Hassabis's statement.
India highlights the disagreement between AI leaders and global tensions, presenting the proposal as part of a struggle for control. The US block on Anthropic's model is highlighted as an example of conflict.
It frames the news as a dispute, using terms like 'do not agree' and 'escalating global tensions' to create a narrative of division.
It omits the details of Hassabis's proposal (such as the FINRA model) and the technical urgency, focusing instead on corporate and government relations.