By Stuart Kerr, Technology Correspondent, LiveAIWire
The AI governance gap became impossible to ignore when OpenAI’s chief scientist Ilya Sutskever, before his departure from the company in 2024, described the trajectory of AI development as entering a phase where the systems being built exceeded the ability of their creators to fully understand or predict their behaviour. This is not a fringe view among AI researchers. The community that has thought most carefully about AI development trajectories is also, disproportionately, the community most concerned about whether the pace at which AI capabilities are advancing is matched by adequate progress in understanding, aligning, and governing the systems being built.
The pace question in AI development is genuinely complex, because pace is not a single variable. The rate at which raw capabilities are advancing is one dimension; the rate at which our understanding of those capabilities, their failure modes, and their social consequences is advancing is another; the rate at which governance frameworks, regulatory requirements, and institutional accountability mechanisms are developing is a third. The concern shared across a wide range of AI researchers, ethicists, and policymakers is that the second and third dimensions are advancing substantially more slowly than the first, creating an AI governance gap that grows with each capability jump.
The Alignment Problem Behind the AI Governance Gap
AI alignment, the challenge of ensuring that AI systems pursue the goals and values their designers intend rather than instrumental objectives that diverge from human wellbeing, has moved from theoretical concern to operational priority for frontier AI laboratories. The emergence of agentic AI systems that plan and execute multi-step tasks without human approval at each stage makes alignment more urgent than it was for simpler AI tools. The techniques developed for aligning simpler systems are being extended to agentic systems but the research community is clear that the safety properties of the simpler systems do not automatically carry over.
The interpretability research needed to understand what happens inside AI models when they process information and generate outputs remains at an early stage relative to the complexity of the systems being studied. Anthropic’s published research on interpretability and alignment is among the most transparent documentation of both progress and remaining challenges in this area, but the gap between current interpretability understanding and the level of understanding needed to verify alignment properties of deployed systems is enormous.
Concentration of Power Widens the AI Governance Gap
One of the most significant dimensions of the AI governance gap is the concentration of frontier AI capability in a small number of organisations. The compute, data, and talent required to develop frontier AI systems creates barriers to entry that mean the most capable AI systems are controlled by a handful of companies primarily located in the United States and, to a lesser extent, China. This concentration has implications for whose values are embedded in the most widely used AI systems and who has meaningful influence over the direction of a technology that is rapidly becoming foundational infrastructure for societies worldwide, a dynamic LiveAIWire has traced in more depth in our coverage of Meta’s superintelligence ambitions.
International governance frameworks for AI, including the Bletchley Declaration and subsequent processes, have begun to address this concentration question but have not produced binding commitments or institutional mechanisms that would meaningfully constrain it. The Seoul AI Summit commitments represent the most recent multilateral effort to build governance norms around frontier AI development.
Autonomous Weapons and the Limits of Existing Law
The military applications of AI represent the dimension of the AI governance gap with the most immediate potential for catastrophic harm. Autonomous weapons systems that can select and engage targets without human decision-making at the point of engagement are being developed by multiple state actors, and the international legal and ethical frameworks governing their use are woefully inadequate. Calls from the International Committee of the Red Cross and a growing number of states for a binding international treaty banning autonomous lethal weapons have not produced agreement among the major military powers investing most heavily in these capabilities, a governance lag LiveAIWire has examined directly in our coverage of AI military strategy.
The Regulatory Divergence Problem
How different jurisdictions are attempting to close the AI governance gap varies substantially, and the divergence itself is a governance challenge. This is examined in detail in LiveAIWire’s coverage of the EU AI Act’s global reach, which documents the most comprehensive legal framework yet attempted for artificial intelligence and the compliance obligations it is already generating well beyond EU borders.
What This Means for You
The AI governance gap is not an abstract philosophical puzzle; it is a question about the governance of technology that is already affecting your working life, the information environment you navigate, the public services you access, and the security environment you live in. Engaging with these questions as a citizen, including through the political and civil society processes that shape AI regulation, and as a user, through the choices you make about which AI products and services you support, is not a specialist activity.
The concentration of decisions about AI development pace and direction in a small number of private companies, whose accountability to the public interest is mediated primarily through market mechanisms and nascent regulatory frameworks, is the meta-ethical challenge underlying the AI governance gap. Changing the outcomes requires changing the incentive structures through regulatory requirements, public funding for safety research, and international coordination that creates accountability beyond what market mechanisms and voluntary commitments can generate. The UK AI Safety Institute’s published work on frontier AI evaluation provides the most transparent public accounting of what is known and not known about the safety properties of the systems currently being deployed.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity, and the social impact of emerging technology. He publishes daily at LiveAIWire.com.