Dominance by understanding is a policy frame named by Kevin Frazier and Alan Z. Rozenshtein in Dominating AI Requires Understanding AI (Lawfare, May 12 2026). It characterizes an apparent third-pole US AI policy stance distinct from both the "doomers" (who favor predeployment government testing mandates) and the "accelerationists" (who favor minimal-intervention dominance per the July 2025 America's AI Action Plan). The central argument is that US AI dominance will require more than faster models; it will require breakthroughs in understanding, testing, and securing frontier AI.
Core argument
Frazier and Rozenshtein argue that understanding a model is a prerequisite for, not an obstacle to, deploying it at scale. They contend the Trump administration has signaled a pivot toward this view, albeit inconsistently and ambiguously, citing a planned executive order to facilitate greater information-sharing between AI labs and government cybersecurity programs; senior-White-House engagement with frontier-lab CEOs in the weeks preceding the article (per WSJ reporting on the Trump-Anthropic-Mythos cluster, see Anthropic); and growing recognition of cybersecurity and biosecurity risk (Economist coverage of bioterrorism uplift; the CISA "CI Fortify" announcement on destructive-cyberattack preparation).
The authors offer three rationales for the pivot. First, the cybersecurity and biosecurity risks of frontier AI are now being taken seriously by White House officials. Second, military and private-sector deployment of the most powerful models will stall until reliability, effectiveness, and controllability are established, an adoption-lag dynamic. Third, technical AI progress and AI dominance are connected, such that understanding the model enables rather than impedes its deployment.
Policy menu
The essay sets out a menu of congressional and executive-branch actions.
The congressional menu has three items: the CREATE AI Act (Young) (NAIRR codification plus a funding increase); the AI Talent Act (H.R. 6573), for federal AI workforce expansion; and NIST appropriations, specifically for the Center for AI Standards and Innovation (NIST CAISI (Center for AI Standards and Innovation)). In support, the article cites a Center for a New American Security (CNAS) gap finding: "For every dollar allocated to NAIRR, the private sector is investing roughly $23,000 in AI." It notes that National Security Commission on Artificial Intelligence (NSCAI) called for $32B in non-defense AI R&D by FY2026, while the realized investment is "about a tenth of that" per Center for Strategic and International Studies (CSIS) analysis.
The executive-branch menu also has three items:
- DPA Section 708 voluntary agreements to coordinate frontier labs on cyber/bio-security mitigation methods and conditional non-deployment agreements, under an antitrust-defense screen administered by the Attorney General and the FTC chair. The Frontier Model Forum is treated as supporting best-practices coordination but unable to lawfully coordinate release decisions without Section 708 antitrust cover.
- CISA workforce restoration. The article notes the administration previously cut roughly 1,000 employees (about one-third of the CISA workforce), that a 75-day DHS shutdown furloughed the remainder, and that a plan to hire 300 mission-critical positions is "a surge in capacity that could not come at a better time." The authors push for hiring beyond 300, alongside use of the Cyber Response and Recovery Fund (6 USC § 677a) for support to critical-infrastructure entities.
- DPA Section 705 quarterly frontier-lab capability survey, confidential and including internal-only deployed models. The article cites the arXiv 2504.12170 capability-disclosure-gap finding to argue labs do not always disclose internal-deployment capabilities to the public.
Distinction from adjacent frames
Frazier and Rozenshtein position the frame as distinct from both poles it sits between. Against predeployment government testing mandates (the "doomer" pole), which were rejected by the May 14-15 industry-coalition pushback from ICBA, BSA, and TechNet per Analysis: Lawmakers, industry pitch frontier AI governance approaches as they await White House moves (Inside AI Policy, May 15 2026), the frame does not require mandatory predeployment vetting; instead it builds the government's understanding capacity alongside lab capacity, with Section 705/708 leverage as the operative instrument. Against minimal-intervention "global dominance" (the "accelerationist" pole), the frame draws on the May 2026 cybersecurity-risk recognition documented in the Frazier-Rozenshtein essay.
The frame is also distinct from, but related to, two further reference points. It differs from Responsible Scaling Policy (RSP) (Anthropic's Responsible Scaling Policy (Version 2.2)): RSPs are voluntary lab commitments to capability-dependent gating, whereas dominance-by-understanding is government-coordinated R&D investment plus DPA leverage. It overlaps with Cyberwar's New Frontier: How AI Agents Will Threaten Global Security (Rosen + Kraprayoon, Foreign Affairs, April 16 2026): Rosen and Kraprayoon's five-part US policy menu (intelligence collection, mandatory incident reporting, CISA restoration, DARPA programs, and KYC for advanced cyber-AI) parallels the Frazier-Rozenshtein menu on CISA restoration but differs on mandatory frontier-lab incident reporting, which Rosen and Kraprayoon favor while Frazier and Rozenshtein favor DPA-mediated voluntary information-sharing.
Relationships
- supports: Dominating AI Requires Understanding AI (Frazier + Rozenshtein, Lawfare, May 12 2026) (canonical source)
- deploys-in: AI and Cybersecurity, AI Pre-Release Vetting (the May 14-15 governance-bidding context), Responsible Scaling Policy (RSP), Defense Production Act (DPA), CREATE AI Act (Young)
- related: Cyberwar's New Frontier: How AI Agents Will Threaten Global Security (Rosen + Kraprayoon, Foreign Affairs, April 16 2026) (parallel five-part US policy menu emphasizing CISA + cyber-AI coordination), Analysis: Lawmakers, industry pitch frontier AI governance approaches as they await White House moves (Inside AI Policy, May 15 2026) (May 14-15 governance-bidding context), Cybersecurity and Infrastructure Security Agency (CISA) — regulator role, National Institute of Standards and Technology (NIST), NIST CAISI (Center for AI Standards and Innovation), Frontier Model Forum, NAIRR — National Artificial Intelligence Research Resource, Anthropic (per the WSJ Mythos-cluster reporting)
- contrasts-with: "Accelerationist global dominance" (America's AI Action Plan, July 2025); "Doomer predeployment testing mandates" (rejected by May 14-15 industry coalitions); Regulatory Capture in AI Policy critique
Sources
- Dominating AI Requires Understanding AI (Frazier + Rozenshtein, Lawfare, May 12 2026) — Frazier + Rozenshtein, Dominating AI Requires Understanding AI, Lawfare, May 12 2026 (canonical anchor)