美国政府迅速将其人工智能监管方法从自由放任的立场转变为高度限制性且不透明的机制。尽管特朗普政府最初保证将避免扼杀创新，但事实上的预先审查制度已经出现。例如，OpenAI的最新模型GPT-5.6 Sol于6月26日被限制在少数“信赖伙伴”中，而Anthropic的Mythos 5在6月30日放宽限制之前，也面临着严厉的出口管制，限制其只能分发给大约100家美国实体及其外籍员工。这种杂乱无章的许可制度创造了一个令人困惑的监管环境，使得像英国人工智能安全研究所这样的国际合作伙伴能够访问某些模型，但被限制访问其他模型。

这种监管转变引发了人们对其将影响美国人工智能实验室竞争力的广泛担忧。开发和训练前沿模型需要巨额的资本投资，实验室通常在发布的最初几个月内收回这些资金，因为此时它们拥有暂时的竞争优势。政府审批造成的延迟蚕食了这一狭窄的窗口，有可能打击未来对数据中心等基础设施的投资。与此同时，中国竞争对手正在迅速追赶，其领先的“开源权重”模型仅落后美国同行六至十个月，且运行成本要低得多。随着美国实验室为了应对政治压力而收紧防护栏，一些公司已经准备绕过受限的美国模型，转而使用中国的开源替代方案。

治理结构本身也面临着重大的执行障碍，其特点是内部技术专业知识严重短缺。虽然白宫计划在8月之前建立一个机密基准测试程序，以评估新模型的网络能力，但私营人工智能实验室中只有极少数人拥有理解这些标准的必要安全许可。此外，像国家网络总监办公室这样的关键决策机构正在经历人员流失，而商务部缺乏裁决复杂安全声明（如模型“越狱”）所需的技术深度。因此，OpenAI和Anthropic都敦促政府建立更具可预测性和透明度的规则，尽管它们在监管应该由联邦机构还是行业资助的自律组织来管理方面存在分歧。

The American government has rapidly transformed its approach to artificial intelligence governance from a libertarian stance to a highly restrictive and opaque regime. Despite initial assurances from the Trump administration that it would avoid suffocating innovation, a de facto preclearance system has emerged. For instance, OpenAI's latest model, GPT-5.6 Sol, was restricted to trusted partners on June 26th, while Anthropic’s Mythos 5 faced stringent export controls that limited its distribution to approximately 100 American entities and their foreign employees before being eased on June 30th. This haphazard system of permissions has created a confusing regulatory environment, leaving international partners like Britain’s AI Security Institute with access to some models but restricted from others.

This regulatory shift has sparked widespread concern regarding its impact on the competitiveness of American AI laboratories. Developing and training frontier models requires massive capital investments that labs typically recover during the initial months of release when they possess a temporary competitive edge. Delays caused by government approvals eat into this narrow window, potentially discouraging future investments in infrastructure like data centers. Meanwhile, Chinese competitors are catching up rapidly, with their leading "open weight" models lagging only six to ten months behind American counterparts while being far cheaper to operate. As American labs tighten guardrails to comply with political pressure, some companies are already preparing to bypass restricted American models in favor of Chinese open-source alternatives.

The governance structure itself faces significant implementation hurdles, characterized by a severe shortage of internal technical expertise. While the White House plans to establish a classified benchmarking process by August to evaluate the cyber capabilities of new models, very few individuals within private AI labs hold the necessary security clearances to understand these criteria. Furthermore, key policy-making bodies like the office of the national cyber director are experiencing staff departures, and the Commerce Department lacks the technical depth required to adjudicate complex safety claims, such as model "jailbreaking." Consequently, both OpenAI and Anthropic are urging the government to establish more predictable, transparent rules, though they disagree on whether oversight should be managed by a federal agency or an industry-funded self-regulatory organization.

Source: Donald Trump’s AI regime is opaque, unpredictable—and unsustainable

Subtitle: The administration now controls who gets access to the best models

Dateline: Jul 02, 2026 08:46 AM