THE SHORT VERSION
- Anthropic wrote to California Governor Gavin Newsom in a letter dated October 15 arguing that when a review finds a model poses a serious risk, the government should be able to ask a court to block it.Anthropic 在 10 月 15 日致加州州长加文·纽森的信中提出，若审查认定模型构成严重风险，政府应有权请求法院阻止该模型。
- The proposal routes independent evaluators' findings to California's attorney general, who could seek a court order pausing a model's release, its internal use or further work on it.该方案让独立评估者向加州总检察长报告风险；总检察长若判断存在近期灾难性危害，可请求法院暂停模型的发布、公司内部使用或后续开发。
- Anthropic also wants outside auditors paid through a state fund and risk reports at least every six months, and it argued that a law focused only on kill switches could give false comfort.Anthropic 还希望外部安全审计员入驻领先 AI 公司、由州设基金付费，并要求大型开发商至少每半年发布风险报告；信中还认为只写「关闭开关」的法律会带来虚假安全感。
A request for a court, not a regulator's order要的是法院，不是监管机构的一纸命令
The letter is dated October 15 and responds to an executive order Newsom signed in September, which set up a working group to recommend how to strengthen California's AI safety law.据 The Next Web 10 月 29 日报道（信件由 Politico 首报），Anthropic 已告知加州州长加文·纽森，州政府应有权请求法院阻止构成严重风险的 AI 模型。该信日期为 10 月 15 日，回应纽森 9 月签署的行政命令；那道命令设立工作组，就如何加强加州 AI 安全法提出建议。
Anthropic's answer is that a shutdown power is useful but insufficient, and that the stronger mechanism is a judicial one.Anthropic 的答案是：关闭权限有用但不够，更有力的机制是司法程序。
From an evaluator's finding to an injunction从评估结论到法院禁令
The mechanism Anthropic describes has three steps. Independent evaluators would report risks to California's attorney general. If the attorney general saw a near-term risk of catastrophic harm, they could ask a court to pause the model. A court order could then stop the model's release, its use inside the company, or further work on it.Anthropic 描述的机制分三步：独立评估者向加州总检察长报告风险；若总检察长判断存在近期灾难性危害，可请求法院暂停该模型；法院命令可阻止模型发布、阻止其在公司内部使用，或阻止继续开发。
That structure keeps a regulator from being the final decision-maker. It also puts a burden of proof in front of a judge rather than in a rulebook: the state would have to persuade a court that the evaluated risk is serious and near-term, and the company would have a forum to contest it. The letter said that when a review finds a model poses a serious risk, the government should be able to ask a court to block it, according to The Next Web.这一结构让监管机构不是最终决策者，也把举证责任放到法官面前，而不是写进规则手册：州政府必须说服法院认定被评估的风险严重且迫在眉睫，企业则获得抗辩的场合。据 The Next Web 报道，信中表示，当审查发现模型构成严重风险时，政府应有权请求法院阻止它。
Auditors inside companies, paid from a state fund审计员进驻公司，费用由州基金支付
The second part of the proposal concerns who does the checking. Anthropic wants outside safety auditors to work inside leading AI companies, with firms paying into a state-run fund rather than paying the auditors directly. That arrangement is meant to remove the obvious conflict that appears when the company being examined also signs the examiner's invoice.方案的第二部分关乎「谁来查」。Anthropic 希望外部安全审计员进驻领先 AI 公司工作，企业向州设基金缴费，而不是直接向审计员付费。这一安排意在消除明显利益冲突：被检查的公司同时为检查者开票。
Anthropic also proposes that large developers publish risk reports at least every six months. Neither idea comes with an enforcement mechanism in the letter as described, and the reporting interval is a floor rather than a schedule tied to releases.Anthropic 还建议大型开发商至少每半年发布风险报告。就报道所描述的内容看，两项建议都没有配套执行机制，报告周期是下限，而不是与模型发布绑定的时间表。
Why Anthropic rejected a kill-switch-only law为什么反对只写「关闭开关」的法律
On kill switches, Anthropic agreed that every leading AI company should be able to shut down its models quickly. It then argued that a law focused only on shutdowns could give false comfort, because the dramatic remedy may not match the actual failure modes operators see day to day.在关闭开关问题上，Anthropic 认同每家领先 AI 公司都应能迅速关停自己的模型。但它随后指出，只聚焦关停的法律可能带来虚假安全感，因为这种戏剧性的救济未必匹配运营者日常面对的真实失效模式。
Instead it backed narrower steps: turning off a single AI agent, or cutting off one customer's access. Those are the interventions that a company running models in production can apply without stopping an entire service, and Anthropic's argument is that they are more likely to be used than an emergency stop.它转而支持更窄的手段：关停单个 AI 智能体，或切断某一位客户的访问权限。这类干预是公司在生产中可执行、又不必停掉整项服务的动作，Anthropic 的论点是：它们比「紧急关停」更可能被真正使用。
OpenAI backs auditors, not court ordersOpenAI 支持审计，不支持法院禁令
Anthropic is not the only company advising the governor. OpenAI sent its own recommendations to Newsom on October 20, backing shutdown controls and in-house auditors but not court orders to block risky models, according to Politico's reporting as summarised by The Next Web. The split is narrow but real: both companies accept external scrutiny, and they disagree on whether a judge should be able to stop a release.Anthropic 并非唯一向州长提建议的公司。据 Politico 报道（The Next Web 转述），OpenAI 于 10 月 20 日提交了自己的建议，支持关闭控制与内部审计员，但不支持以法院命令阻止风险模型。分歧不大却很实在：两家公司都接受外部审视，分歧在于法官是否应能阻止一次模型发布。
Anthropic is making a similar case in Washington. A draft federal bill it has circulated would let the Commerce Department ask a federal court to block a model's release in narrow cases, and House lawmakers have proposed an AI Kill Switch Act that would let officials order companies to shut down dangerous models. California's working group is the state-level version of the same argument.Anthropic 在华盛顿也在推类似主张。它流传的一份联邦法案草案，将允许商务部在狭窄情形下请求联邦法院阻止模型发布；众议院议员还提出《AI 紧急关闭法案》，授权官员命令企业关停危险模型。加州工作组是同一条论证在州层面的版本。
What the letter settles and what it does not这封信确定了什么，又没有确定什么
Nothing has been decided. The letter is advice to a working group set up by the September order, and no bill, rule or court ruling has followed it. Newsom's office had not responded when the reporting was published, so the governor's position is unknown.目前没有任何决定落地。这封信是 9 月行政命令所设工作组的建议；此后没有法案、规则或法院裁决跟进。报道发布时纽森办公室尚未回应，州长立场未知。
The letter also does not define what counts as a serious risk or a near-term catastrophic harm, which are the terms the whole mechanism turns on. Those definitions would have to be written into law before a court could apply them. Related coverage on this site includes the House bill setting a 48-hour clock for AI incident reports, Anthropic's three-tier cyber verification programme and the DOJ's questions about Nvidia's Hugging Face deal. Each is a different lever — disclosure, access control, competition review — and none of them is a court injunction over a model release.信中也未界定何为「严重风险」、何为「近期灾难性危害」——而整套机制的支点正是这两个词。它们必须先写入法律，法院才可能适用。本站相关报道包括 众议院法案为 AI 事件报告设定 48 小时时限、Anthropic 三层网络验证计划与美国司法部就英伟达与 Hugging Face 交易提出疑问。三者分别是披露、访问控制与竞争审查三种杠杆，都不是针对模型发布的法院禁令。
A company asking for more scrutiny of itself主动要求更多监督的公司
Anthropic's case is unusual because it asks for more outside scrutiny of the systems the company itself builds. The letter proposes a court that could block a model and auditors paid from a state fund, and the company says both are needed because a shutdown power alone is too blunt. A lab arguing for that kind of oversight is making a narrower claim than the headline suggests. The same letter asks for risk reports at least every six months and for outside auditors paid from a state-run fund.Anthropic 的论证有一处特殊背景：它请求对自家构建的系统施加更多外部审查。信中建议设立可阻止模型的法院程序，并由州基金支付外部审计员费用；公司称两者都必要，因为只有关停权限太生硬。一家主动要求这类监督的实验室，其主张比标题看起来更窄。同一封信还要求至少每六个月发布一次风险报告，并要求由州设基金支付外部审计员的费用。
That history explains the shape of the proposal. Anthropic is not asking for a regulator to be able to switch a model off at will; it is asking for a court, a standard of near-term catastrophic harm and a defined set of remedies. Whether that is stronger or weaker than a kill-switch law depends on how the terms are written, and the letter, as reported, does not write them.这段历史解释了提案的形态：Anthropic 要求的不是监管机构可随意关停模型，而是法院、以「近期灾难性危害」为标准，以及一组明确界定的救济方式。这比关停法案更强还是更弱，取决于条款如何书写，而据报该信并未写出这些条款。
Sources & reporting limits来源与核查范围
Follow Policy & safety to find related stories on your next visit.关注政策与安全，下次访问时阅读相关文章。
