微软总裁布拉德·史密斯表示，目前人工智能所受到的监管远远不够，甚至不如牛奶和洗衣机等普通消费品。他呼吁各国尽快建立全球性的人工智能安全规则，包括要求企业在人工智能系统失控并造成损害时通知受害者、提交事故报告，以及在部分人工智能产品正式发布前接受独立评估机构的审查。
史密斯的这一立场与美国总统唐纳德·特朗普的主张形成了明显对比。特朗普认为，人工智能行业应当主要依靠自身进行监管，避免过多政府干预削弱美国企业与中国竞争对手较量的能力。史密斯则认为，人工智能已经同时具备工具和武器的双重属性，不能继续仅仅依靠企业自律来控制风险。
史密斯在接受英国广播公司（BBC）《Big Boss Interview》播客采访时表示：“现在，你家的洗衣机受到的安全规则都比人工智能更多。你在杂货店看到的每一盒牛奶，都受到安全规则和检验的约束。”
他认为，人工智能的发展已经到了必须采取行动的阶段，各国应当借鉴其他行业的监管经验，建立适用于人工智能技术的安全制度，而不是等到严重事故发生后才开始讨论如何应对。
近年来，随着人工智能模型的能力迅速提升，要求对这一行业实施更严格监管的声音不断增加。一些人工智能领域的专家甚至警告，这项技术可能对人类社会的长期安全构成威胁。围绕人工智能可能带来的网络安全风险、失控行为和其他潜在危害，行业内部也开始出现关于是否应该放慢技术开发速度的讨论。
美国人工智能企业Anthropic、OpenAI以及GoogleDeepMind的负责人都曾支持对人工智能发展采取更加谨慎的态度，部分人士还支持在特定情况下放缓技术研发。不过，特朗普一直反对可能限制美国人工智能产业发展的措施，并多次强调，美国必须在这场技术竞争中保持领先。
史密斯则认为，人工智能不应被简单地视为一种普通的软件产品。随着人工智能代理系统获得更强的自主执行能力，以及先进模型在网络安全领域的影响日益扩大，相关风险已经需要政府、企业和独立评估机构共同应对。
他呼吁企业为人工智能系统设置紧急制动机制，也就是业界常说的“终止开关”（Kill Switch）。如果人工智能系统出现失控行为，或者开始执行可能造成严重后果的操作，企业应当能够及时暂停系统运行，或者降低其运行速度和权限，防止风险进一步扩大。
史密斯还主张建立强制性的事故报告制度。如果人工智能代理失控并侵入其他系统，或者因异常行为对用户造成损害，相关企业就应当依法报告事故，并告知受到影响的个人或组织。
他认为，这些措施应当成为人工智能安全监管体系的一部分，而不能仅仅依赖企业自愿采取行动。
此外，史密斯建议，在某些高风险人工智能产品发布之前，应由政府认可的独立评估人员进行审查，判断系统是否符合相关安全要求。这并不意味着所有人工智能产品都必须接受相同程度的监管，而是应当根据产品能力、使用场景及潜在风险，决定是否需要在发布前进行独立评估。
史密斯表示，其他行业已经建立了类似的安全管理机制，人工智能监管可以借鉴这些成熟经验。他认为，关键在于把已有的风险评估、事故报告和产品安全制度，合理地应用于人工智能领域。
在推动国际人工智能规则方面，史密斯特别看好英国的作用。
他指出，英国已经建立了专门研究人工智能安全问题的机构。该机构最初在前首相里希·苏纳克执政期间以人工智能安全研究机构的形式推出，后来更名为人工智能安全研究所（AI Security Institute），负责研究先进人工智能系统的能力与潜在风险。
史密斯认为，英国在人工智能安全评估方面具有较强的技术实力，也具备推动国际合作的条件。他表示，如果询问科技行业人士，哪个国家在人工智能安全技术能力方面处于领先地位，很多人会毫不犹豫地选择英国。
他认为，英国有机会利用自身的研究能力和国际影响力，推动各国围绕人工智能安全建立更加统一的规则。
史密斯的观点与特朗普政府强调加快人工智能发展的政策方向存在差异。特朗普多次表示，谁赢得人工智能竞争，谁就将赢得未来。这一立场反映出美国政府希望通过减少监管障碍、鼓励企业投资和扩大技术应用，巩固美国在人工智能领域的领先地位。
不过，史密斯并不认为美国政府一定会拒绝由英国主导的人工智能监管倡议。
他指出，美国政界看待英国与欧盟的态度存在差异。在华盛顿，英国和伦敦通常不会像布鲁塞尔及欧盟机构那样，引发同等程度的政治抵触情绪。史密斯认为，这种关系上的差异可能成为英国推动国际人工智能安全标准的一项优势。
他强调，当前已经出现了足够多值得重视的风险信号，包括可能脱离人类控制的人工智能代理，以及能力强大到足以对全球网络安全构成威胁的模型。
史密斯表示：“当你看到脱离人类控制的人工智能代理，当你看到强大到足以危及全球网络安全的模型时，我会说，时机已经成熟。”
他呼吁各方采取冷静但坚定的态度，尽快展开对话，推动政府、企业和研究机构共同制定相关规则。他还表示，希望微软能够在这一过程中发挥作用，也希望英国能够承担推动国际合作的角色。
值得注意的是，史密斯并没有完全否定行业自律的价值。上月底，特朗普曾在白宫会见OpenAI、Anthropic、英伟达、SpaceX、Meta和Google等科技巨头的负责人。与会企业高管签署了一份具有道德约束性质的文件，同意与独立审计人员合作，评估人工智能系统是否按照预期运行，并确保系统不会以非预期方式入侵或访问技术系统。
这一文件旨在推动人工智能企业加强安全审查，并通过独立评估发现潜在问题。尽管其约束力与正式法律法规不同，但仍被视为科技行业在安全治理方面采取的一项行动。
史密斯对这份声明表示赞赏，并称微软愿意签署相关文件。他认为，这是一个良好的开端，也是必要的一步。
但在他看来，仅仅依靠这一类自愿承诺已经不够。
史密斯指出，任何自我监管制度都会面临一个根本性问题：负责任的企业通常愿意遵守安全要求，但真正需要担忧的，是那些缺乏责任意识、不愿主动采取安全措施的公司。
他认为，如果整个行业仅仅依靠自愿承诺，那么遵守规则的企业可能会承担更多成本，而不愿投入资源保障安全的竞争者却可能继续快速扩张。因此，建立具有法律约束力的最低安全标准，有助于确保所有企业都必须达到基本要求。
这也正是史密斯所强调的核心问题：人工智能监管不应仅仅取决于企业是否愿意承担责任，而应当建立明确的制度，规定企业必须履行哪些安全义务，以及在系统出现严重问题时需要采取什么行动。
随着人工智能代理逐渐获得执行复杂任务、调用外部工具和访问技术系统的能力，传统的软件安全管理方式可能需要作出调整。未来的监管制度不仅需要关注模型本身是否存在缺陷，还需要考虑人工智能在实际运行中能够获得哪些权限、可以访问哪些资源，以及发生异常行为时是否能够及时中止操作。
对于微软而言，这一问题也具有直接的商业意义。微软既是人工智能基础设施和模型应用的重要参与者，也是企业级云计算和软件服务的主要供应商。人工智能系统的安全性、可靠性和可控性，将直接影响企业客户是否愿意在业务流程中广泛部署相关技术。
史密斯的表态表明，即使在美国科技行业内部，对于人工智能应当如何监管，也并不存在完全一致的看法。一部分企业和政策制定者希望尽可能减少限制，以加快创新和商业化进程；另一部分声音则认为，随着技术能力增强，必须同步建立更明确的安全责任和风险控制机制。
这场争论的关键，已经不仅是是否支持人工智能发展，而是如何在鼓励创新的同时，确保技术应用不会对个人、企业和社会造成无法接受的风险。
从牛奶和洗衣机的例子，到紧急制动机制、事故报告和独立评估，史密斯试图传达的信息十分明确：人工智能的技术发展速度正在超过现有监管体系的调整速度。如果各国希望继续从人工智能中获得经济和社会收益，就需要尽快建立与其能力和风险相匹配的安全规则。
而在他看来，单纯寄希望于企业自律并不足以解决这一问题。真正有效的监管体系，必须既能鼓励负责任的企业继续创新，也能约束那些不愿主动控制风险的参与者。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Mr. Brad Smith, President of Microsoft, said that artificial intelligence was currently under far less regulation than ordinary consumer goods such as milk and washing machines. He called on States to establish, as soon as possible, global rules for the safety of artificial intelligence, including the requirement for enterprises to notify victims when artificial intelligence systems were out of control and caused damage, to submit accident reports and to be reviewed by independent assessment bodies prior to formal publish for some artificial intelligence products.
This position of Smith is in clear contrast to that of United States President Donald Trump. According to Trump, the artificial intelligence industry should rely primarily on its own regulation to avoid excessive government intervention that weakens the ability of United States firms to compete with Chinese competitors. Smith, for his part, believes that artificial intelligence already has the dual attributes of both tools and weapons and cannot continue to rely solely on corporate self-regulation to control risk.
In an interview with BBC 's Big Boss Interview, Smith said: "Now, your washing machine has more safety rules than artificial intelligence. Every box of milk you see in a grocery store is subject to safety rules and tests."
In his view, the development of artificial intelligence had reached a stage where action had to be taken, and States should build on the regulatory experience of other industries and establish safety systems applicable to artificial intelligence technology, rather than wait until a serious accident had occurred before beginning to discuss how to respond.
In recent years, there has been a growing demand for stricter regulation of the industry, as the capabilities of artificial intelligence models have increased rapidly. Some experts in the field of artificial intelligence have even warned that this technology may pose a threat to the long-term security of human society. Discussions have also begun within industry on whether technology development should be slowed down around the potential network security risks, uncontrolled behaviour and other potential hazards posed by artificial intelligence.
The heads of the United States artificial intelligence firms Anthropic, Open AI and Google DeepMind have all supported a more cautious approach to artificial intelligence development, and some support slowing technology development in specific contexts. Trump, however, has been opposed to measures that could restrict the development of the United States artificial intelligence industry and has repeatedly stressed that the United States must remain in the lead in this technological competition.
Smith argued that artificial intelligence should not simply be seen as an ordinary software product. As the artificial intelligence agent system acquires greater self-implementing capacity and advanced models have grown in influence in the area of cybersecurity, the associated risks already require a combination of government, business and independent evaluation agencies.
He called on enterprises to put in place an emergency brake mechanism for artificial intelligence systems, known as “terminating the switch” (Kill Switch), as the industry often called it. If an artificial intelligence system is out of control or starts operations that may cause serious consequences, the enterprise should be able to suspend the system in a timely manner or to reduce its speed and permission to prevent further risk.
Smith also advocated a mandatory accident reporting system. If artificial intelligence agents are out of control and hacking into other systems, or if they cause damage to the user as a result of an unusual act, the enterprise concerned shall, in accordance with the law, report the accident and inform the affected person or organization.
In his view, such measures should be part of an artificially intelligent safety regulatory system and not rely solely on voluntary action by enterprises.
In addition, Smith suggested that prior to certain high-risk artificial intelligence products, publish should be reviewed by independent government-accredited evaluators to determine whether the system met the relevant safety requirements. This does not mean that all artificially intelligent products must be subject to the same level of regulation, but rather should decide whether an independent assessment before publish is required, depending on product capacity, the context of use and potential risks.
Smith indicated that similar safety management mechanisms had been established in other industries and that artificial intelligence regulation could build on those mature experiences. In his view, it was essential to rationally apply existing systems of risk assessment, accident reporting and product safety to artificial intelligence.
Smith particularly values the role of Britain in promoting international artificial intelligence rules.
He noted that the United Kingdom had established a body specializing in artificial intelligence safety. It was first launched under the former Prime Minister, Rich Sunak, in the form of an artificial intelligence security research institute, and was later renamed the AI Security Intelligence Institute, which is responsible for studying the capabilities and potential risks of advanced artificial intelligence systems.
Smith believes that Britain has a strong technical capacity in artificial intelligence security assessment and is equipped to promote international cooperation. He said that many would not hesitate to choose the United Kingdom if they asked the technologists which country was in the lead in artificial intelligence safety technology.
He believed that the United Kingdom had the opportunity to use its research capacity and international influence to promote more uniform rules around artificial intelligence security.
There is a difference between the views of Smith and the policy direction of the Trump Government, which emphasizes the acceleration of artificial intelligence. Trump has repeatedly stated that whoever wins the competition for artificial intelligence will win the future. This position reflects the desire of the United States Government to consolidate its leadership in the area of artificial intelligence by reducing regulatory barriers, encouraging business investment and expanding technology applications.
However, Smith did not believe that the United States Government would certainly reject the British-led Artificial Intelligence Regulatory Initiative.
He noted that there was a difference between the attitude of United States politics towards Britain and the European Union. In Washington, Britain and London usually do not generate the same degree of political resistance as Brussels and European Union institutions. Smith argued that such differences in relationships could be an advantage for Britain in promoting international artificial intelligence safety standards.
He stressed that there were already enough risk signals to warrant attention, including artificial intelligence agents that could be removed from human control, and models that were strong enough to pose a threat to global cybersecurity.
Smith said, “When you see artificial intelligence agents out of human control, when you see models that are strong enough to endanger global cybersecurity, I say, the time is ripe.”
He called on all parties to take a calm but firm attitude and to engage in a dialogue as soon as possible to promote joint rule-making among Governments, enterprises and research institutions. He also expressed the hope that Microsoft would play a role in that process and that the United Kingdom would assume a role in promoting international cooperation.
It is noteworthy that Smith did not completely negate the value of industry self-regulation. At the end of last month, Trump met with the heads of the technological giants OpenAi, Anthropic, Weeda, SpaceX, Meta and Google at the White House. The senior management of the participating enterprises signed an ethically binding document agreeing to work with independent auditors to assess whether the AIIS is functioning as expected and to ensure that the system does not invade or access technical systems in an unexpected manner.
This document aims to promote enhanced safety vetting by artificially intelligent enterprises and to identify potential problems through independent assessments. Although its binding force is different from that of formal laws and regulations, it is still seen as an action taken by the scientific and technological industry in the area of security governance.
Smith expressed appreciation for the statement and stated that Microsoft was willing to sign the relevant documents. In his view, that was a good start and a necessary step.
In his view, however, such voluntary commitments alone were no longer sufficient.
Smith points out that any self-regulation system faces a fundamental problem: Responsible businesses are usually willing to comply with safety requirements, but what is really a matter of concern is those companies that lack a sense of responsibility and are reluctant to take safety measures on their own initiative.
In his view, if the industry as a whole relied solely on voluntary commitments, compliance firms might bear additional costs, while competitors who did not want to invest resources to guarantee security might continue to expand rapidly. Therefore, the establishment of legally binding minimum safety standards helps to ensure that all enterprises must meet basic requirements.
This is the core issue highlighted by Smith: Artificial intelligence regulation should not depend solely on the willingness of the enterprise to assume its responsibilities, but should establish a clear system of what safety obligations the enterprise has to meet and what actions are required in case of serious systemic problems.
The traditional software security management approach may need to be adjusted as artificial intelligence agents gradually acquire the capacity to perform complex tasks, access external tools and access technical systems. Future regulatory regimes need to look not only at whether the model itself is flawed, but also at what permission is available to artificial intelligence in practice, what resources are accessible and whether the operation can be suspended in a timely manner in the event of an anomaly.
This issue also has direct commercial significance for Microsoft. Microsoft is both an important participant in the application of artificial intelligence infrastructure and models and a major provider of enterprise-level cloud computing and software services. The security, reliability and controlability of artificial intelligence systems will have a direct impact on the willingness of corporate clients to deploy relevant technology widely in business processes.
Smith ' s statement suggests that there is no complete agreement on how artificial intelligence should be regulated, even within the United States science and technology industry. Some businesses and policymakers want to minimize constraints in order to accelerate innovation and commercialization, while others argue that, as technological capacity increases, clearer safety responsibilities and risk control mechanisms need to be established simultaneously.
The key to this debate is not just whether to support artificial intelligence development, but how to encourage innovation while ensuring that technology applications do not pose unacceptable risks to individuals, enterprises and societies.
From the case of milk and washing machines to emergency brake mechanisms, accident reporting and independent evaluation, Smith tried to convey a clear message that artificial intelligence technology is developing faster than the adjustment of existing regulatory systems. If States wish to continue to derive economic and social benefits from artificial intelligence, they need to establish security rules that match their capabilities and risks as soon as possible.
In his view, it was not enough to rely solely on corporate self-regulation. A truly effective regulatory system must both encourage responsible enterprises to continue to innovate and constrain those who are unwilling to proactively control risks.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.