Microsoft’s Brad Smith Says the UK Can Lead Global AI Safety Regulation

Preface

As artificial intelligence advances, questions about cybersecurity, accountability, and public safety are growing more urgent. Microsoft President Brad Smith has called for stronger oversight, arguing that products such as milk and washing machines face more regulation than the companies developing AI. Speaking on a BBC podcast, he described AI as both a “weapon and a tool” and said the time has come for global safety rules. Smith believes the United Kingdom is especially well placed to help lead that effort. Its AI Security Institute, launched in 2023, has access to leading models before deployment and conducts safety testing. This article reviews Smith’s argument, the UK’s potential role, recent industry commitments, and the regulatory measures he says could help address the risks posed by increasingly capable AI systems.

Lazy bag

Smith’s central argument is that AI safety cannot depend solely on companies choosing to regulate themselves. He sees the UK as a credible convenor because its AI Security Institute tests leading models and, in his view, Britain can engage with AI oversight in ways that Washington may find politically difficult. Recent industry pledges to work with independent auditors are a constructive step, but Smith says they do not replace binding rules. His proposed safeguards include reporting breaches involving rogue AI agents, independent evaluation before release, and mechanisms to pause or slow systems if people lose control.

Main Body

Microsoft President Brad Smith is urging governments to establish stronger safeguards for artificial intelligence. His case begins with what he sees as a regulatory imbalance: everyday consumer goods, including milk and washing machines, are subject to oversight, while the companies creating powerful AI systems operate without comparable requirements. Smith argues that this gap deserves urgent attention as AI becomes more capable and increasingly connected to sensitive digital systems.

In a recent BBC podcast interview, Smith characterized AI as both a “weapon and a tool.” The description captures the dual nature of the technology: AI can support useful services and innovation, but it can also contribute to cybersecurity threats and other forms of harm. Smith warned that AI is putting the world’s cybersecurity at risk and said the “moment has arrived” to establish global safety rules. His remarks reflect a broader concern that technical capabilities are moving faster than the frameworks intended to govern their development and use.

Smith argues that the United Kingdom, rather than the United States or the European Union, is best positioned to help lead international AI safety efforts. A central reason is the UK’s AI Security Institute, which was launched in 2023. The institute receives access to major models from U.S. frontier labs before those models are deployed, allowing it to test their safety and capabilities. This access has helped the institute become a prominent participant in AI safety evaluation.

The institute’s role has also been affected by political tensions. U.S. President Donald Trump has warned against what he calls “globalist” organizations, which he says interfere with American companies. Last month, Trump reportedly asked Anthropic and OpenAI to withhold new models from British testers. However, the AI Security Institute confirmed on Monday that it is able to access and test the latest models. This episode illustrates how international cooperation on AI oversight can be influenced by government relationships as well as technical considerations.

Despite the political resistance, Smith believes the UK may be able to approach regulation in ways that are harder for the United States. He told the BBC, “In Washington, D.C., people don’t look at the U.K. and they don’t look at London the way they look at Brussels.” He described that distinction as a strength. His point is that Britain may be able to convene stakeholders and advance safety discussions without being viewed through the same political lens as the European Union’s regulatory institutions.

Smith’s comments come amid calls from technology leaders to take the pace and consequences of AI development seriously. Anthropic’s Dario Amodei and OpenAI’s Sam Altman have both called for slowing the pace of development, following several incidents in which AI agents went rogue. Such events have intensified concerns about systems acting in unexpected ways, particularly when they are given access to tools, data, or computer environments. The incidents reinforce the importance of testing how AI behaves beyond ideal or controlled conditions.

Last month, Trump hosted technology leaders from OpenAI, Anthropic, Nvidia, AMD, SpaceX, Meta, and Google at the White House to discuss AI safety. The participants signed a “morally binding” document in which they agreed to work with “independent auditors” to assess the safety and capabilities of their AI models. The pledge represents a commitment to external review, but it is not the same as a legally enforceable regulatory framework.

Smith described the document as a “good step,” while emphasizing that it falls short of the regulation he believes is necessary. He highlighted a weakness in voluntary self-regulation: responsible companies may choose to comply, while less responsible organizations may not. As he put it, “It’s the companies that are not as responsible that you need to worry about.” In Smith’s view, rules that rely only on the willingness of each company to participate may leave gaps precisely where oversight is most needed.

Among the safeguards Smith would like to see is a legal obligation for an organization to notify an appropriate body if its systems are breached by rogue AI agents. A reporting requirement could help authorities and other affected organizations understand incidents and respond to emerging threats. It would also establish that serious AI-related security events are not solely internal matters for the companies involved.

Smith also appears to support outside evaluators reviewing AI systems before products are released. Independent assessment could provide an additional check on a developer’s own testing and help identify safety or capability concerns before deployment. Such a process would not eliminate uncertainty, but it could make evaluations more transparent and ensure that release decisions are informed by scrutiny beyond the organizations that built the systems.

Another measure he discussed positively is the use of “kill-switches” or comparable mechanisms that can pause or slow an AI system if humans lose control of it. The idea is to preserve a practical means of intervention when a system behaves unexpectedly. For these safeguards to be useful, organizations would need to consider in advance who can activate them, under what circumstances, and how the system can be safely paused. Smith’s remarks present such controls as part of a wider approach rather than a substitute for regulation.

Ultimately, Smith called for a measured but firm effort to bring people together. “Let’s speak calmly, maybe even softly, but firmly, and start to bring people together,” he said. He expressed hope that Microsoft could contribute to that process and that the UK could do so as a country. His position combines international cooperation with enforceable safeguards: independent testing and industry participation can contribute to safety, but, in his view, they should sit alongside clear legal duties and mechanisms for intervention.

The debate described by Smith is therefore about more than which country writes the rules. It concerns how governments, AI developers, independent evaluators, and security organizations can coordinate as models become more capable. The UK’s testing role offers one possible foundation for collaboration, while the disagreement over voluntary commitments shows why policymakers continue to consider binding requirements. Smith’s argument is that rules should apply not only to companies willing to cooperate, but also provide meaningful protection when they are not.

Key Insights Table