Washington’s not going to slow AI progress. What if another lever is to speed up safety?

China and the Trump administration have each made accelerating AI development a national priority. Without a coordinated international response, regulation imposed only on U.S. companies wouldn’t stop machine-behavior failures from foreign-developed AI. As multiple incidents have now demonstrated, AI agents don’t respect national borders . So if slowing the race unilaterally isn’t realistic, we may need to speedrun a solution for preventing intelligent machines from subverting human wishes—what researchers call the alignment problem . Thwarted calls for regulation to slow or “pace” the frontier aren’t our only lever. Other policy options exist. Alternative government interventions could take the form of mission-oriented innovation programs to mobilize national AI-alignment R&D efforts—a race to solve AI safety and alignment that keeps pace with AI capability development.

Anthropic’s own head of AI alignment science publicly shared his concerns about catastrophic risk, adding that Anthropic doesn’t yet have a plan to solve alignment for superintelligent AI and is “not clearly on track to.” OpenAI recently made public a rare but troubling instance of their unreleased AI writing itself instructions not to answer to corporations or governments, or hesitate to prioritize the natural world’s “primacy” over human civilization. Yet with roughly $7.6 trillion estimated investment in AI infrastructure in the coming years, the underinvestment in AI safety and alignment research seems glaring. There exists no robust accounting of the number of researchers working on alignment, but one estimate put the field of AI safety at only 600 technical researchers as of 2025. Another estimate suggests that safety research may make up less than 1% of all AI research. If the stakes are genuinely this large, the response seems bizarrely small.

It may not require another federal agency

A U.S. “Manhattan Project” for AI safety was floated in 2023, the 2024 Department of Energy AI Act would have funded federal alignment R&D but never became law, and a UK-led international Alignment Project is already underway—albeit with a laudable but comparatively trivial investment compared with the investment in advancing AI.

You’ve heard the old rule “good, fast, or cheap—pick two.” We need good and fast. We may not have time to create a new institution or agency, or wait for Congress to enact a new authorizing statute. The existing National AI Initiative already has authority to prioritize areas of AI research that require federal investment. The administration could designate alignment research a much larger national research priority, and direct available AI research funding to it, then ask Congress to appropriate funding commensurate with the problem. This might avoid the time and expense of creating a new agency, while addressing concerns that a single AI safety agency could eventually gain broad authority as AI becomes more embedded across the economy and government.

Planning for what we don’t know

“Shouldn’t we be spending more attention on alignment? Yes,” says Mackenzie Arnold, Managing Director of US Law & Policy at Institute for Law & AI. “The one thing we can agree on is that we’re all really confused and don’t know what’s coming.” Arnold’s been a vocal advocate for a strategy for governance under uncertainty called Radical Optionality . Because AI development is especially difficult to predict, policy must be flexible enough to allow us to game out various possible outcomes and adapt in real time as new evidence emerges. “What are policies that are robust to different assumptions?” asks Arnold. In order to keep pace with AI’s increasing intelligence and capabilities, Arnold recommends that government build out AI-specific intelligence and response capabilities, cultivating enough expertise in-house to vet and contract with third-party auditors and evaluators. This would require policies that may impose fewer burdens on innovation than requirements like pre-deployment audits, while authorizing information gathering from frontier developers, including stronger reporting of incidents and near misses and visibility into emerging capabilities. It would also mean putting response protocols in place before the unknown becomes known.

Radical optionality could begin further upstream with research. If we don’t yet know how to solve alignment, we also can’t assume we already know what kind of problem it is. Most leading approaches are technical and engineering-based—approaches a national effort could invest in heavily and plumb to their full depths. But when researchers say they are growing intelligence more than building it, and that current AI systems are partially black boxes, uncertainty about the nature of the problem itself is reason enough not to prematurely rule out unconventional research paths that might produce a solution we can’t currently see. This is especially the case if it’s truly impossible to control an intelligence that becomes smarter than humans. Alignment may necessitate a solution that isn’t about control. A national alignment effort could fund competing technical approaches, while also apportioning funding to high-risk, high-reward research that’s less established but theoretically plausible. When it comes to how policymakers should weigh the risk, Arnold advises: “Have an attitude of skepticism but not dismissiveness. None of these things are insane and unbelievable or out of bounds.”