Article reader Listen + reading controls
Article reader
Preparing the reader…
Reading settings
A Successful AI Pilot Is Not Yet a Durable Institution¶
Revised and substantially expanded July 17, 2026, including the subsequent legislative and program record.
In February 2024, lawmakers supporting the Creating Resources for Every American To Experiment with Artificial Intelligence Act—the CREATE AI Act—were optimistic that Congress would formally establish the National Artificial Intelligence Research Resource (NAIRR). At the same time, the National Science Foundation had already launched a pilot under executive direction and existing authorities.
That combination created an easy but misleading story: the pilot was underway, bipartisan support was growing, and legislation would simply make the program permanent.
The harder institutional question is what permanent should mean.
A pilot can demonstrate demand, test operating processes, and mobilize voluntary contributions. It cannot by itself guarantee multi-year infrastructure, stable access for long-running research, independent governance, predictable appropriations, enforceable duties, or continuity across administrations. Those are not administrative details surrounding a technical program. They determine what kind of national capability the program becomes.
NAIRR needs legislation not merely so that it continues to exist, but so that its public obligations survive the incentives and relationships that made the pilot possible.
The Historical Record Is More Instructive Than the 2024 Optimism¶
The original FedScoop report captured genuine bipartisan momentum. The 118th Congress did advance the proposal: the House Science, Space, and Technology Committee ordered H.R. 5077 reported, as amended, by voice vote in September 2024, and a Senate version was reported with an amendment in December. But the legislation did not become law before that Congress ended.
In March 2025, lawmakers introduced H.R. 2385, the CREATE AI Act of 2025. As of this July 2026 revision, Congress.gov lists the bill as introduced and referred to the House Committee on Science, Space, and Technology. Meanwhile, the pilot continued to grow. NSF's two-year update reports more than 600 research and education projects and 6,000 students across all states, the District of Columbia, and Puerto Rico.
This is not a story of failure. It is a case study in the gap between program momentum and institutional settlement. The pilot can become operationally important before Congress resolves its durable governance and funding model. Indeed, that very success increases the cost of ambiguity: more researchers, institutions, and projects begin to depend on a capability whose future terms are not fully settled in statute.
What Legislation Can Do That a Pilot Cannot¶
Good legislation should resist the temptation to freeze a rapidly changing technical architecture. Congress does not need to specify accelerator models, cloud services, foundation-model families, or orchestration tools. It should establish the durable mission, authorities, public protections, and accountability mechanisms within which technical choices can evolve.
At least seven functions require statutory clarity.
1. Establish a public mission and eligibility boundary¶
NAIRR exists to improve the nation's capacity for AI research, education, testing, and discovery—not to provide an undifferentiated subsidy for commercial product development. The statute should define eligible people and institutions, permissible purposes, and the public-interest outcomes expected from the portfolio.
The 2025 bill would make U.S.-based researchers, educators, students, agencies, and federally funded research and development centers eligible under defined conditions. That is an essential starting point. Implementation should still distinguish between basic access, public-interest research, independent evaluation, workforce development, and commercially adjacent work so each can be governed appropriately.
2. Create accountable decision rights¶
A federated resource needs a clear principal. Federal agencies contribute assets; private firms donate services; an operating entity may run day-to-day functions; advisory bodies provide expertise; and researchers consume resources. Without explicit decision rights, responsibility can diffuse across the coalition.
Statute should identify who sets portfolio priorities, selects and oversees the operating entity, admits resources, allocates scarce capacity, manages incidents, resolves conflicts, evaluates performance, and answers to Congress and the public. Delegation is necessary. Accountability must remain traceable.
3. Provide stable, flexible funding¶
Compute allocations and donated services can be episodic, while research and workforce development are longitudinal. Projects need confidence that environments, identities, data, and artifacts will remain available long enough to complete and reproduce work. The operating organization needs technical staff and support capacity that cannot be rebuilt annually.
Multi-year authorization and appropriations should support a diversified portfolio: public infrastructure, acquired commercial services, grants or allocations, secure research environments, data stewardship, evaluation capabilities, training, and operational staff. A resource supported almost entirely by in-kind commercial contributions would be inexpensive in budgetary terms but fragile in strategic terms.
4. Protect independent and reproducible research¶
NAIRR will be valuable partly because it can enable research about systems operated by powerful organizations. Researchers must be able to publish credible positive or negative findings, subject to legitimate security, privacy, intellectual-property, and coordinated-disclosure constraints.
Legislation or implementing policy should require transparent resource terms, preserve academic freedom, document provider influence, protect appropriate criticism, and ensure access decisions are not used to suppress disfavored results. Reproducibility also requires model and service version records, exportable experiment artifacts, and notice when providers make material changes.
5. Govern security without making it a universal barrier¶
The resource will serve projects with very different risk profiles. A classroom exercise using open data should not inherit the controls of a project using sensitive health or critical-infrastructure data. Conversely, a generic account and terms-of-service agreement cannot protect a high-consequence research environment.
NAIRR needs tiered security, privacy, identity, data, and research-security controls based on resource and use-case risk. The 2025 bill's minimum security requirements are important; they should become a floor within a risk-based architecture, not a one-size-fits-all ceiling.
6. Require a continuity and exit architecture¶
National research infrastructure must survive vendor departures, service retirements, funding interruptions, operating-entity transitions, and changes in technical fashion. Continuity plans should cover researcher identities, project records, datasets, model and experiment artifacts, security evidence, publications, and active allocations.
Every proprietary contribution should have an exit plan before it becomes a critical dependency. The operating entity itself should be replaceable through documented interfaces and transferable records. Institutional permanence should not mean permanent vendor lock-in.
7. Make evaluation a statutory habit¶
Congress should require independent, recurring evaluation of access, additionality, research outcomes, safety, concentration, portability, operational performance, and workforce impact. Reporting only aggregate usage will reward activity instead of public value.
The evaluation should include unsuccessful applicants, projects that failed to onboard, institutions that received resources but could not convert them into results, and research domains the portfolio did not reach. Those negative spaces reveal whether the infrastructure is changing the research ecosystem or merely serving institutions already equipped to succeed.
Avoid Two Opposite Legislative Errors¶
Congress faces two common failure modes.
The first is under-specification: authorize a broad program, celebrate public-private collaboration, and leave difficult questions of independence, allocation, data stewardship, security, and continuity to informal arrangements. This maximizes near-term flexibility but allows path dependence to become policy.
The second is technical over-specification: encode current architectures, thresholds, or categories so precisely that the statute becomes obsolete as AI changes. A provision designed around today's model access patterns may constrain tomorrow's distributed, embodied, scientific, or edge-AI research.
The better pattern is to legislate durable properties and require adaptive implementation. Those properties include:
- public-interest purpose;
- equitable and transparent access;
- independent research and publication;
- risk-based security and privacy;
- interoperability and portability;
- data and model provenance;
- competitive, conflict-aware acquisition and allocation;
- continuous evaluation;
- operational and financial continuity;
- public reporting with appropriate protection for sensitive information.
Agencies and the operating entity can then update standards, interfaces, and controls through transparent processes as technology evolves.
“Democratization” Requires More Than Geographic Reach¶
NAIRR is often justified as a way to democratize AI research. That term should be operationalized rather than repeated.
Geographic participation and institutional diversity are important, but access is not democratized if researchers outside elite centers receive accounts yet lack the engineering support to use them. Nor is it democratized if the only practical pathway trains researchers into proprietary ecosystems whose future costs their institutions cannot afford.
A serious democratization measure would examine whether NAIRR changes:
- who originates and leads advanced AI research;
- which questions receive resources;
- who can independently evaluate dominant models;
- which institutions retain durable technical capability after an award;
- whether students gain transferable rather than provider-specific skills;
- whether research artifacts remain available to the broader community;
- whether underserved communities participate in setting priorities, not only supplying data or serving as study populations.
These outcomes require technical assistance, research engineering, curriculum, mentorship, accessible proposal processes, and sustained institutional partnerships. Compute is necessary. It is not sufficient.
The Pilot Should Produce Legislative Evidence¶
The pilot's greatest near-term contribution may be evidence about how to write and implement the permanent program. NSF and its partners should make the program's operational learning legible to Congress:
- demand by resource type and research domain;
- time from application to productive work;
- allocation fairness and reviewer consistency;
- provider and platform concentration;
- steady-state costs after contributions;
- security and data-governance incidents;
- user-support burden and skill gaps;
- portability and reproducibility outcomes;
- reasons projects fail, withdraw, or underuse allocations;
- public outputs and additional research made possible.
This evidence would allow lawmakers to authorize a capability grounded in actual operating experience rather than hypothetical architecture. It would also help appropriators distinguish unavoidable cost from avoidable inefficiency.
The Strategic Case Is About National Optionality¶
The United States benefits from leading AI companies. It also benefits from universities, laboratories, nonprofits, startups, agencies, and independent researchers able to pursue questions that those companies cannot or will not prioritize. NAIRR's strategic function is to preserve that pluralism.
National optionality means the country can investigate an emerging risk without waiting for a platform owner, develop a method whose commercial market is uncertain, train researchers outside a few firms, compare competing technical approaches, and sustain scientific artifacts longer than a product cycle. It is a form of resilience in the national knowledge base.
That is why a successful pilot is not enough. Pilots are optimized to demonstrate motion. Institutions are designed to sustain purpose through changes in leadership, budgets, vendors, and technology.
The legislative question is not simply whether Congress likes NAIRR. It is whether Congress will define and fund the conditions under which NAIRR remains a genuinely national research resource after the novelty of the pilot and generosity of early contributors fade.
Questions of durable AI institutions, technical governance, and public-sector operating models are central to my professional and research portfolio. If you are turning a promising AI pilot into a capability that must survive real organizational constraints, you can connect with me on LinkedIn or send a direct inquiry.