Christoph Winter and Charlie Bullock, writing for the Institute for Law & AI, begin with a dilemma anyone thinking seriously about AI governance must confront. Transformative AI, which they define as AI capable of precipitating a transition comparable to or greater than the agricultural or industrial revolution, may arrive within a decade. But the right regulatory response cannot be known in advance. Move too early with rigid rules, and governments risk stifling innovation, targeting the wrong risks, or building legal regimes that technological progress overtakes before they take effect. Wait too long, and they may lack the institutional capacity to act when action becomes unavoidable.
Their paper, Radical Optionality: Governing Transformative AI Under Uncertainty, flagged by Jack Clark in Import AI 456, rests on premises they argue are reasonable enough to compel action: transformative AI may arrive soon; its capabilities, risks, and appropriate governance responses remain profoundly uncertain; a dual-use technology with national security implications will require some government oversight; and the capacity to govern it will take years to build.
The structural imbalance behind these premises is the "pacing problem," in which technological progress routinely outstrips the legal, regulatory, and institutional capacity required to govern it. AI widens that gap. Its systems are difficult to understand; they can perform at expert level in one setting and fail spectacularly in another; and recursive improvement could make capability progress faster than institutions can adapt.
Winter and Bullock warn that if progress becomes exponential, human beings are "psychologically disinclined to accept the implications of exponential growth." This makes the challenge even harder for governance because its absence is less visible than capability progress. Governance capacity has to be built before the gap becomes obvious.
Against this backdrop, Winter and Bullock propose a third path that rejects both premature heavy regulation and passive inaction. They call it radical optionality: aggressive, proactive investment in the information, expertise, and institutional capacity governments will need to respond competently and flexibly to a wide range of futures, including ones no one has anticipated. The paper is careful, politically realistic, and honest about what it does not know. It treats uncertainty as a reason to prepare, not a reason to wait.
Better information, deeper expertise, and more flexible rules are necessary. They still leave the hardest question unresolved: how can governance remain answerable to independent human judgment at the speed and scale of AI itself? Winter and Bullock's policies build the right inputs. The solution lies in organizing those inputs into shared operating infrastructure.
What Governance Infrastructure Means
AI is now embedded in finance, logistics, military, public services, and critical infrastructure operations, where it shapes decisions and workflows rather than sitting at the edge as a feature. Political and industrial actors increasingly frame AI as "essential national infrastructure," akin to electricity or roads, with multi-layer stacks spanning energy, compute, data centers, models, and applications. That framing reflects capital allocation and policy attention, not just rhetoric.
In my own work, I have come to see the tension between security and innovation as primarily an infrastructure problem, with policy playing an important but contextual role. If AI operates as infrastructure, governance cannot arrive only as a patchwork of policy. It has to meet infrastructure with infrastructure. Governance infrastructure is still a forming category, while AI capability advances exponentially beyond the systems built to govern it.
“If AI operates as infrastructure, governance cannot arrive only as a patchwork of policy. It has to meet infrastructure with infrastructure.”
At a high level, governance infrastructure gives governance the operating capacity to function as a persistent environment rather than a sequence of bounded events. Its purpose is to give governance the scale, speed, and elasticity proportionate to the systems it is meant to govern.
Governance infrastructure gives operational form to the information, expertise, and institutional capacity Winter and Bullock call for. It brings evidence, standards, and system behavior into a computational system built to produce outputs. The actors may include labs, deployers, regulators, standards bodies, evaluators, users, or affected communities. Those outputs can persist, interoperate, and update as systems change, making governance usable across deployment contexts at the pace AI demands.
Such infrastructure processes governance-relevant evidence and the signals AI systems generate in operation. It transforms those inputs into analytical outputs that clarify descriptive problems and generative outputs that create new governance instruments where existing behavioral metrics are inadequate or missing. Both directly support radical optionality's goal of building institutional capacity before critical decision points arrive.
From Inputs To Capacity
Winter and Bullock's evaluation pillar shows why inputs alone are not enough. A vibrant third-party ecosystem, including organizations like METR and Apollo Research, can produce findings about frontier models that governments, labs, and the public would not otherwise have. But findings become governance only when they can be interpreted, acted on, and revised over time. What does a result mean? How should it change system behavior? Who must agree? How should that agreement change as the system changes? These are governance questions, not evaluation outputs.
The same is true of the competing perspectives governance must engage. Labs, regulators, deployers, users, and affected communities see risk differently and respond to different incentives. Those differences are themselves a source of risk information. A useful governance architecture connects that information to system behavior: how incentives shape what systems are built, disclosed, deployed, or contested, and how system behavior changes what principals know, demand, or withhold. That connection turns risk data into new capacity for governments deciding when and how to act, and for builders shaping what comes next.
This is where public and private roles should be complementary. Winter and Bullock are right about what distinguishes private governance from public authority: the former cannot supply the state's coercive power or its formal role as coordinator among affected actors. Privately built governance infrastructure can contribute a different advantage: faster innovation, focus, and specialized talent directed at methods and tools public institutions are structurally unlikely to develop quickly enough on their own. The near-term role for government is to build institutional capacity while supporting those private-sector initiatives that complement its role as convener for the actors whose decisions shape how AI systems are built, deployed, evaluated, and used.
The Investment
The United States built the Interstate Highway System over roughly 35 years, at a cost equivalent to about $20 billion annually in today's dollars. Research estimates that removing the system would reduce real GDP by more than $600 billion every year, a recurring annual loss that exceeds the system's entire one-time construction cost. The construction cost was paid once. The system became a permanent condition of American economic scale.
But that scale depended on the system being usable at acceptable risk. Bridges, grading, load ratings, guardrails, signage, and other operating conditions were not afterthoughts. No one proposed building the highways first and adding the conditions for safe use later. Without those conditions, there would have been no national freight network at comparable scale, no equivalent expansion of commerce, and no highway system capable of supporting America's postwar economic expansion.
Radical optionality will require radical investment. AI capex reaches into the hundreds of billions annually and is projected to reach $1.6 trillion annually by 2031. Investment in governance infrastructure is a rounding error by comparison. That level of spending carries its own premise. If capital is being deployed as if AI is infrastructure, the implicit claim is that it will be used at scale, reliably, and by institutions that trust it. Governance is the condition that makes that implicit claim true. Infrastructure built for scale and speed ultimately fails without the conditions that make it governable and trusted.
The deeper lesson is that governance remains too external to the AI stack. Safety and security should become operating conditions of AI deployment itself, reducing the pressure to choose between speed and control. That will not happen by modestly extending the existing conversation. It will require treating governance infrastructure as a place for invention, and funding the unorthodox ideas most likely to make that invention possible.
“It will require treating governance infrastructure as a place for invention, and funding the unorthodox ideas most likely to make that invention possible.”
Governments and funders that stop at evaluation capacity will have done something valuable. Those that back the inventors of governance infrastructure may help build something decisive.