Dean Ball, the primary architect of America's AI Action Plan, announced he is joining OpenAI to build a new team focused on frontier AI policy. Ball spent the last year and a half at the White House Office of Science and Technology Policy, where he drafted the government's comprehensive strategy for artificial intelligence development and deployment. His decision to move into the private sector reflects a growing recognition that the most consequential decisions about advanced AI systems are increasingly made inside labs, not in regulatory frameworks.

What Is the AI Action Plan, and What Did It Miss?

Ball describes the AI Action Plan as a "hermeneutics exercise," a document written for today's Washington audience that he hopes a more AI-literate future version of those same officials will re-read and finally understand. The plan reads more like three dozen thematic objectives than one cohesive strategy, according to Ball's assessment. Several adoption case studies he was passionate about, including hospital record-keeping systems and the Veterans Affairs system as a single-payer data goldmine, were left on the cutting-room floor due to time constraints.

Ball

On implementation, Ball estimates he is roughly 30 to 40 percent done on his mental to-do list after eleven months. He points to real wins on nuclear energy, Federal Energy Regulatory Commission (FERC) grid-interconnection reform, and military adoption. However, he warns that senior officials have departed from the plan's original spirit, most notably by imposing global export controls on frontier models "with about 90 minutes' notice," confirming the exact fear foreign partners had voiced to him: that the United States would restrict access to advanced AI systems when political tensions rise.

Why Is Ball Concerned About Government Monopolies on Frontier AI?

Much of Ball's recent work has focused on a troubling trend: the concentration of frontier AI decision-making inside government agencies. He revisits the moment the Department of War, the renamed Department of Defense, declared Anthropic a "supply chain risk," a designation now being litigated before the D.C. Circuit. Even as other agencies continue using Anthropic's models, the company is being wound down inside the Department of War.

Ball notes that the National Security Agency (NSA), technically part of the Department of War, not only maintains its Anthropic contract but has honored Anthropic's red lines around domestic mass surveillance and autonomously lethal weapons. His deeper concern centers on the cyber executive order's voluntary 30-day pre-deployment testing program, whose details are classified and run primarily by the intelligence community. This creates a future in which models the public doesn't know exist are tested against standards that cannot be disclosed.

Ball frames this problem through a systems-level lens: a society is an "information processing system," with all its humans functioning as parallel compute working out how to handle each new level of AI capability. Centralizing those decisions inside a small, low-context circle of officials makes policymaking brittle and prone to errors that ripple across the entire economy.

How Are States Reshaping AI Governance?

Ball is bullish on states as laboratories for AI policy. He highlights the converging transparency provisions of California's SB 53, New York's RAISE Act, and Illinois' SB 315, along with the spread of independent verification organizations. These state-level efforts create a more distributed approach to AI oversight. However, he flags messier state laws, like Illinois' ban on AI mental-health services, as examples of where real patchwork problems emerge.

• California SB 53: Transparency requirements for AI systems used in high-stakes decisions
• New York RAISE Act: Algorithmic accountability and impact assessment mandates
• Illinois SB 315: Disclosure and consent requirements for AI-driven decisions
• Independent Verification: Third-party organizations emerging to audit AI systems across jurisdictions

What Surprised Ball About AI's Technical Trajectory?

Ball is candid that little about the technical trajectory has thrown him. Scary cyber capabilities arrived roughly on schedule. However, two developments did surprise him: the runaway popularity of coding agents, particularly among homeschooling communities using Claude Code and OpenClaw, and world-simulation models suddenly gaining object permanence. The latter development collapsed his timeline for solving dexterous robotic manipulation to "about eight months".

Ball experienced a personal disappointment with what he calls the "Fable ban." He had access to Anthropic's Fable model only briefly, long enough to watch it demolish a 70-page expert rebuttal in his own FERC Order 1000 proceeding, before it was pulled. This marked the first time, he notes, that users have ever gone backwards in capability: "your AI literally became dumber in the last week." Ball's read on the government's reaction blends genuine security concern, a lack of frontier-AI context, and undeniable political coloring. The official story eventually landed on Anthropic's expanded-access rollout reaching SK Telecom, a South Korean partner that Ball argues is hardening-worthy, not an obvious threat.

Why Is Ball Joining OpenAI Now?

Ball explains that he ultimately concluded he couldn't think clearly about the frontier lab as a new "center of political and economic power" without being inside one. His historical analogy is the birth of modern finance in the Dutch Republic and Britain. His OpenAI team won't function as a traditional lobbying shop, which is Chris Lehane's Global Affairs group. Instead, it will look six to twelve months ahead, work closely with technical staff on where capabilities are heading, and grapple with the reality that the most consequential decisions around internal deployments and recursive self-improvement will increasingly be made before any public release, outside the reach of regulations triggered by deployment.

Ball

Ball is careful to say he retains an independent public writing voice. What he is ultimately after, he explains, is getting "this whole transformation right for the country and the world," invoking his self-description as a civis Americanus, a citizen of America.

How Should AI Labs and Governments Balance Power?

On the question of recursive self-improvement (RSI), Ball's prior leans continuous rather than discontinuous, but he advocates to "measure twice and cut once." He is intrigued by mechanisms like an Federal Trade Commission (FTC) "no-action letter" that would publicly signal that narrowly scoped coordination on a slowdown wouldn't be treated as cartel behavior. However, he warns that anti-competitive moves dressed up as safety, citing Anthropic's Fable output-degradation "safeguards," undermine the whole case and would have to be carefully scoped around.

On the labs' leverage against a state that holds the monopoly on violence, Ball's answer is twofold: the genuine national-security capability the models unlock, such as National Geospatial-Intelligence Agency (NGA) data overhangs and Project Maven shrinking missile-targeting teams from 2,000 to 20, and broad diffusion as a political check. He channels Charles Tilly's concept of the interdependence of states and capital. Ball concedes the government could, in principle, invoke Defense Production Act "priorities authority" to commandeer compute, an idea echoing Leopold Aschenbrenner's line about stitching together the nation's data centers in a crisis. But he argues it won't, because the state and the labs are interdependent, and because broad diffusion gives every bank, university, and industry a stake that makes confiscatory outcomes far less likely.

"He wants AI treated not as a specific industry, but as capital itself," Ball explained in the conversation.

Dean Ball, Senior Fellow at the Foundation for American Innovation, now joining OpenAI

Ball's move to OpenAI marks a significant moment in how frontier AI labs are organizing their relationship with government. Rather than waiting for regulation to arrive, OpenAI is building internal capacity to anticipate policy questions and shape the conversation around how advanced AI systems should be developed and deployed. This approach reflects a broader recognition that the most important decisions about AI's future are being made not in Congress or regulatory agencies, but in the labs themselves, months or years before any public announcement.