AI governance is moving from promises to proof

Over the past three years, AI governance debates have primarily focused on issues like the impact on jobs, copyright, data privacy, and liability. However, a recent UK parliamentary debate suggests that we may be witnessing a shift in the focus of AI policy. Over 70 MPs and peers have called for legislation prohibiting the development of artificial superintelligence (ASI) and an international agreement to prevent its creation.

While this may not become government policy immediately, the debate signals a broader move from technological policy to national security considerations.

This shift is significant because it implies that policymakers are more willing to impose restrictions on AI development, even if it comes with economic costs. Companies accustomed to dealing with technology ministries and regulators will now encounter defense establishments, security agencies, and heads of government. This transition mirrors the experiences of telecommunications infrastructure and semiconductor supply chains, which have already undergone a similar evolution.

Moreover, industry leaders are increasingly acknowledging the need for pacing AI development. Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman have publicly supported the idea of "pacing the frontier" and implementing independent safety assessments. Even Elon Musk, a competitor to both companies, has endorsed Amodei's proposal.

This represents a significant change in the regulatory debate, as the question may no longer be simply whether governments should constrain AI development, but rather how fast the frontier should move and what safeguards must accompany each increase in capability.

If AI governance evolves into a verification process akin to financial regulation, with independent testing, incident reporting, and measurable thresholds, we could see profound changes for businesses. Amodei has proposed giving independent external evaluators ongoing access to Anthropic's operations, drawing an analogy to regulatory supervisors in financial institutions.

OpenAI has now committed to doing the same. This move could compel companies to demonstrate their commitment to safety through demonstrable safeguards, rather than simply making promises.

For Britain, this presents a delicate balancing act. The government seeks to position the country as a major AI power, attracting investment, talent, and development of increasingly capable models. At the same time, it must reassure voters that these technologies will not pose unacceptable risks. This challenge is further complicated by the geopolitical nature of AI development.

If democratic countries slow their programs while competitors do not, AI could create a national-security vulnerability. Therefore, international coordination becomes increasingly important.

Rather than seeking a global prohibition on superintelligence, governments may find more practical ground for cooperation. This could involve prohibiting narrowly defined dangerous applications, developing common testing standards for cyber and biological risks, or establishing internationally recognized capability thresholds. As policymakers begin to discuss AI in the conceptual language of arms control, rather than digital regulation, the nature of AI governance is undergoing a fundamental shift.