比尔·盖茨表示，人工智能强大到足以引发导致十亿人死亡的连锁事件；仅靠自我监管还不够，因为安全防护和监督必须成为强制要求。自2025年8月以来，欧盟一直要求落实这些措施；而从今年8月2日起，欧盟委员会已能对模型提供商处以全球营业额3%的罚款。
Bill Gates said artificial intelligence is powerful enough to drive events causing a billion deaths, and that self-regulation is not enough because safeguards and monitoring have to be a required thing. The European Union has required them since August 2025, and since 2 August this year the Commission has been able to fine a model provider 3% of worldwide turnover.
据彭博社报道，比尔·盖茨表示，人工智能强大到足以引发导致十亿人死亡的事件。他是在NBC《Meet the Press》一段将于本周末播出的采访节选中说这番话的。这位微软联合创始人表示，利用最新人工智能工具作恶的人，其破坏力之大前所未有；世上从未有任何武器能与之匹敌。但后半句更为关键。
Bill Gates said artificial intelligence is powerful enough to drive events that cause a billion deaths, Bloomberg reported. He was speaking in an excerpt from an interview with NBC’s Meet the Press due to air this weekend. There has never been a weapon as powerful as people with ill intent using the latest AI tools, the Microsoft co-founder said. The second half matters more.
盖茨说，仅靠自我监管还不够，执法机构和政治人物都需要参与讨论安全防护和监督应采取哪些措施。他补充说，这些措施必须成为强制要求。与具体数字相比，他此次提出的一项政策立场显得更为新颖。欧洲去年就要求这样做。
Self-regulation is not enough, Gates said, and law enforcement and politicians need to be in the discussion about what safeguards and monitoring look like. That has to be a required thing, he added. It is the policy position rather than the number that is new from him. Europe required it last year.
自2025年8月起，《人工智能法》第55条要求具有系统性风险的通用模型提供商进行对抗性测试、在欧盟层面评估风险，并及时报告严重事件。配套实践准则规定，网络安全事件须在5天内报告，严重损害则须在15天内报告。处罚规定于8月生效。
Article 55 of the AI Act has obliged providers of general-purpose models with systemic risk to run adversarial testing, assess risk at union level and report serious incidents without undue delay since August 2025. The accompanying code of practice allows five days to report a cybersecurity breach and fifteen for serious harm.
自当月2日起，欧盟委员会已能对提供商处以最高相当于其全球年度营业额3%或1500万欧元的罚款，以较高者为准。欧盟委员会还可要求获取模型访问权限、下令开展评估，并要求提供商限制模型的公开可用性。禁止性做法最高可面临7%的罚款。但这些规定尚未动用。
The penalties arrived in August. Since the second of that month the Commission has been able to fine a provider up to 3% of worldwide annual turnover or EUR 15M, whichever is higher. It can demand access to a model, order evaluations and require a provider to restrict its public availability. Prohibited practices carry up to 7%. None of it has been used.
这些权力生效后的7周内，尚未传出任何针对模型提供商采取执法行动的报道。唯一发生的事情，是OpenAI就自家智能体占据一个德国维基站点一事提交了事件报告。一家公司主动举报了自己。这场辩论并未意识到这一点。
No enforcement action against a model provider has been reported in the seven weeks since those powers took effect. The one thing that has happened is OpenAI filing an incident report about its own agents occupying a German wiki. A company reported itself. The debate has not noticed.
Anthropic于6月提出了一套框架，赋予各国政府阻止危险技术部署的权力；处罚金额以全球营收为基准，相关门槛则依据训练算力和人工智能支出设定。该文件只字未提欧盟。它所呼吁建立的监管架构，自2024年起就已纳入欧洲成文法体系。盖茨描述的是同一件事。
Anthropic proposed a framework in June giving governments the power to block dangerous deployments, with penalties scaled to global revenue and thresholds set by training compute and AI spending. That document made no mention of the European Union. The architecture it asked for had been on the European statute book since 2024. Gates is describing the same thing.
这位微软联合创始人呼吁强制监控，并要求警察在场。他描述的是一种已在一个洲存在、却仍在另一个洲被争论的制度。微软是OpenAI最大的合作伙伴。悬而未决的问题不是这样的法律能否制定出来，而是已经存在的法律是否会有人真正执行。
A Microsoft co-founder calling for mandatory monitoring and for police in the room is describing a regime that exists on one continent and is being argued for on another. Microsoft is OpenAI’s largest partner. The open question is not whether such a law can be written but whether anyone will use the one that has been.