When AI-driven science outpaces its institutions

The discussion between Peter Diamandis and Michael Kratsios, director of the White House Office of Science and Technology Policy (OSTP), in the 04 August 2026 Moonshots episode, American Science Policy and AI-Driven Research Reform, is often framed as a productivity question. AI systems can search literatures, generate hypotheses, write code, interpret large datasets, and increasingly participate in experimental workflows. The prospect is not only faster science. It is a different relationship between researchers, research organizations, and the technical systems through which knowledge is produced.

My current book project, When Policy Falls Behind: Bitcoin, AI, and the Governance of Fast Systems, begins from a related problem. Institutions do not respond to technical change at a uniform speed. Formal rules may change quickly when a problem becomes politically visible, while professional norms, organizational routines, funding arrangements, and accumulated ideas about legitimate expertise change slowly. Douglass North's framework gives this asymmetry much of its force. Institutions persist because people have learned their rules, built organizations around them, developed reputations within them, and formed expectations about how others will behave.

AI-driven research places that settlement under pressure.

Research institutions are often described as cumbersome because they contain review processes, funding cycles, publication delays, professional boundaries, and layers of oversight. Some of that friction is plainly wasteful. Yet science is not organized only to produce more results. Its institutions also determine who can challenge a result, what counts as adequate evidence, how error is corrected, and which problems receive sustained public attention. A system that produces candidate findings quickly still requires ways to establish whether those findings travel beyond the particular model, dataset, laboratory, or commercial relationship that produced them.

The difficulty appears when the rate of technical change exceeds the rate at which those institutional functions can be revised. AI can widen the gap between recognition and response. A research funder may recognize that automated systems are altering a field while lacking the expertise, information, or authority needed to adjust its evaluation criteria. A university may adopt AI tools before it has settled questions about authorship, data access, replication, or responsibility. A public agency may support AI-enabled discovery while depending on privately controlled models, compute, and experimental data that it cannot independently inspect.

These are governance questions rather than secondary implementation details. Control over a model or dataset can become an institutional position when other researchers cannot readily reproduce the work without access to the same technical infrastructure. The issue is not that private firms should be excluded from scientific research. Their capabilities may be indispensable in many areas. The issue is whether public institutions retain enough standing and practical capacity to evaluate results, redirect priorities, and identify where commercial incentives are narrowing the research agenda.

AI also changes the economics of scientific attention. Where generating plausible hypotheses becomes cheaper, the scarce resource shifts toward judgment: deciding which questions deserve experimental capacity, which results require independent testing, and which apparent advances have enough public value to justify institutional commitment. That judgment cannot be delegated entirely to the same systems that generate the possibilities. A technical system may rank outputs efficiently according to its objectives while remaining unable to determine what there is most reason to want from a public research system.

The policy task is therefore more demanding than accelerating procurement or expanding compute access. American science policy needs institutions that can support faster experimentation while preserving independent evaluation, credible correction, and public priority-setting. That will require arrangements for access to data and models, records that make AI-assisted work legible, and evaluation processes that can respond more quickly without becoming merely performative. It will also require clearer distinctions between scientific capacity that is publicly governed and capacity that is available only through particular firms or platforms.

I do not treat institutional latency as an argument for slowing technical development - it is a diagnostic condition. The relevant question is whether the governance form surrounding AI-enabled research fits the speed, concentration, and uncertainty of the activity it is meant to govern. Where it does not, more scientific output may coexist with weaker public capacity to decide what that output means, who benefits from it, and how its direction can be changed.

That is the broader concern behind When Policy Falls Behind. Fast systems create pressures that cannot be resolved by asking institutions to work harder at their existing pace. They require arrangements that preserve recognition, judgment, and meaningful response after the conditions of action have already changed.