布莱切利园 AI 安全峰会

2023 年 11 月 1 日到 2 日，英国在布莱切利园召开了首届 AI 安全峰会。这个地点有特殊的象征意义：二战期间，盟军在这里破译了德军的恩尼格玛密码。选在这里开会，英国政府显然是想要传递一个信号——AI 安全是一个需要全球协作应对的重大问题。来自中国、美国、欧盟等地的政府代表，与 OpenAI、DeepMind、Anthropic 等前沿实验室齐聚一堂。

峰会最重要的产出，是各国签署的《布莱切利宣言》。宣言的措辞经过反复博弈，最终达成的共识是：前沿 AI 可能带来「重大乃至灾难性」的风险，包括网络攻击、生物武器扩散等滥用场景，以及未来更强大模型带来的失控风险。这是第一次在政府间层面，各国正式承认前沿 AI 存在灾难性风险，并承诺开展国际合作。

当然，宣言本身没有法律约束力。它更像一份「问题确认书」——各国承认问题存在、同意继续对话，但具体怎么治理、由谁监管、边界在哪，全部留给后续。批评者指出，峰会缺少具体的监管框架和可执行承诺，更像一场「高规格的研讨会」。但支持者认为，把 AI 安全抬到国家元首级议程这件事本身，就是历史性的第一步。

峰会开启了一个持续进行的系列。2024 年首尔峰会、2025 年巴黎峰会相继召开，讨论从「承认风险」逐渐走向「如何测量风险」和「如何评估前沿模型」。各国也陆续推出自己的监管动作——欧盟 AI 法案、美国的行政命令、中国的生成式 AI 管理办法。布莱切利园打开的那扇门，被后来的会议不断推开。

对产业界来说，峰会的直接影响是让安全评估成为「行业标准动作」。前沿实验室开始制度化地进行红队测试、风险分级、能力评估，并把结果披露给政府。开发者也在模型发布流程里把安全评测列为常规环节。一场没有强制力的会议，通过设定议题的方式，改变了整个行业的工作习惯。

回看布莱切利峰会，它的价值在于「第一次」。第一次，AI 安全从学界和实验室的内部讨论，变成了多国政府共同参与的国际议程；第一次，「前沿模型」和「灾难性风险」成为全球治理的标准话语。它没有解决任何具体问题，但它让全世界开始认真讨论这些问题——在 AI 历史的时间线上，这是一道清晰的分界。

On November 1-2, 2023 the UK hosted the first AI Safety Summit at Bletchley Park. The venue carried special symbolism: during WWII, Allied codebreakers cracked the German Enigma cipher there. Choosing it, the UK government clearly meant to signal that AI safety is a major problem requiring global cooperation. Government representatives from China, the US, Europe, and elsewhere gathered with frontier labs including OpenAI, DeepMind, and Anthropic.

The summit's most important output was the Bletchley Declaration signed by participating countries. Its wording was hammered out through intense negotiation; the resulting consensus was that frontier AI may pose "serious, even catastrophic" risks—including misuse scenarios like cyberattacks and bioweapon proliferation, as well as runaway risk from more powerful future models. This was the first time at an intergovernmental level that countries formally acknowledged catastrophic risk from frontier AI and pledged international cooperation.

Of course, the declaration has no binding legal force. It reads more like a "problem acknowledgment document"—countries agree the problem exists and will keep talking, but specifics of governance, who regulates, and where boundaries lie are all left to follow-ups. Critics noted the summit lacked concrete regulatory frameworks and enforceable commitments, resembling a "high-level seminar." But supporters argued that lifting AI safety onto heads-of-government agenda was itself a historic first step.

The summit opened a continuing series. Seoul in 2024 and Paris in 2025 followed, with discussions moving from "acknowledging risk" toward "how to measure risk" and "how to evaluate frontier models." Countries also issued their own regulatory moves—the EU AI Act, US executive orders, China's generative-AI measures. The door Bletchley opened was pushed wider at every subsequent meeting.

For industry, the summit's direct effect was making safety evaluation a "standard industry move." Frontier labs began institutionalizing red-teaming, risk tiering, and capability evaluation, disclosing results to governments. Developers also made safety evaluation a routine step in model release pipelines. A meeting without enforcement power changed industry working habits by setting the agenda.

Looking back at the Bletchley summit, its historical position lies in "first time." For the first time, AI safety moved from academic and lab-internal discussion to an international agenda joined by multiple governments; for the first time, "frontier models" and "catastrophic risk" became standard vocabulary of global governance. It solved no specific problem, but it made the world begin seriously discussing these problems—a clear dividing line on AI's timeline.