在历史上，危机往往催生新的监管机制，如1907年股市暴跌一半后创立的美联储，以及大萧条时期成立的证券交易委员会。虽然人工智能尚未造成重大灾难，但诸如Anthropic的Mythos和OpenAI的GPT-5.6 Sol等前沿模型已具备极强的黑客能力，并可能协助生物恐怖主义。为此，特朗普政府正试图在灾难发生前实施监管，但其目前的尝试显得非常混乱。例如，在Anthropic发布限制版模型Fable后，政府于6月中旬对其施加出口管制，却又在6月30日匆忙解除禁令，这凸显了政府决策的随意性和地缘政治色彩。

对前沿AI进行永久性封锁在技术上是行不通的。中国顶尖模型如Z.ai发布的GLM 5.2已经赶上了美国上一代最先进的水平，且大多数是开源的，任何人都可以运行和修改。尽管由于芯片短缺和美国限制知识蒸馏，中国实验室追赶Mythos等模型需要更多时间，但这种限制最多只能争取几个月或一年的时间。此外，美国科研人员自身也高度依赖便宜且易于修改的中国模型。将前沿AI模型长期封锁还会剥夺公众逐步适应技术的机会，一旦未来突然释放，将会对社会造成更剧烈的冲击。

合理的监管路径应当是建立技术评估机制和面向可信机构的分阶段释放机制。虽然确定风险容忍度是政治决策，但具体的技术评估应交给专业机构。美国的人工智能安全研究所和英国的人工智能安全倡议（AISI）等机构可以与私营部门合作，借鉴金融和电力行业的联合监管模式，整合行业与学术界的专业知识。面对美国可能长期出于国家利益而滥用出口管制和限制措施的风险，其他国家必须通过扶持本土AI产业、建设非美数据中心和转向非美模型来建立自身的应对和防御机制。

Historically, major crises have driven regulatory creation, such as the panic of 1907 leading to the Federal Reserve. While artificial intelligence has not yet caused a disaster, models like Anthropic's Mythos and OpenAI's GPT-5.6 Sol possess advanced hacking and potential bioterrorism assistance capabilities. In response, the Trump administration is attempting to regulate the industry, but its actions have been chaotic. For example, it imposed export controls on Anthropic's guardrailed model, Fable, in mid-June, only to abruptly lift the ban on June 30th. This intervention highlighted the government's willingness to make up rules on the fly with a nationalistic bias favoring American access over foreign users.

A permanent blockade on frontier AI is both unworkable and undesirable. Chinese AI models, such as Z.ai's open-weight GLM 5.2, already match the best of the previous generation of US models. Although chip shortages and crackdowns on model distillation may delay Chinese labs, they remain only months to a year behind. Furthermore, many American researchers rely on cheap, malleable Chinese models. Restricting access to advanced models also prevents society from adapting gradually to technological leaps. Bottling up multiple generations of advances would concentrate massive power among a selected few, and their eventual release would unleash sudden, disruptive chaos.

To safely release advanced models, governments should establish formal evaluation periods and staggered releases to trusted institutions. While elected leaders must set risk tolerance limits, they should avoid micromanaging technical assessments. Specialized bodies, like America's Centre for AI Standards and Innovation and Britain's AI Safety Institute, should oversee evaluations, potentially adopting joint-governance frameworks from finance and power sectors. Given the strategic risks of US export controls, foreign countries must develop leverage by supporting domestic AI sectors and ensuring businesses can transition to non-American models and data centers.

Source: America should not imprison frontier AI

Subtitle: Fable is free. But the technology desperately needs better regulations

Dateline: Jul 02, 2026 08:46 AM