AI Safety Advocacy Splits on US-China Cooperation vs. Domestic Controls

The AI safety community is divided between two governance strategies: the AI Futures Project's 'Plan A,' which calls for a temporary US-China cooperative pause on frontier AI development while existing capabilities drive economic growth [1], and a domestic-controls approach centered on export restrictions and executive authority [3][2]. The Trump administration has declined to build a formal AI licensing regime, leaving governance to ad hoc executive discretion [4]. Anthropic's Pentagon partnership collapsed after the Defense Department demanded blanket usage rights that overrode Anthropic's stated limits on mass surveillance and autonomous weapons [4].

Whether AI safety is treated as a cooperative global problem or a unilateral national security matter has concrete consequences: a cooperative pause requires Chinese participation that is politically contested, while purely domestic export controls risk pushing capability development into less safety-conscious hands. The Anthropic-Pentagon breakdown shows these are not abstract debates—redlines either hold or they don't when a major contract is on the table.

Two governance visions are competing within the AI safety community. The AI Futures Project's 'Plan A' argues that the US and China should agree to a temporary pause on frontier AI development while existing capabilities—already sufficient to drive substantial economic growth given adequate data-center build-out—are deployed and safety infrastructure is constructed [1]. Proponents like Daniel Kokotajlo frame this not as anti-growth but as a prerequisite for capturing AI's benefits: 'If you want to actually benefit from that growth and abundance, you need to avoid the risks involved' [1]. The cooperative approach faces obvious practical obstacles, including the question of verification and what one account describes as 'pretty far-out ways of ensuring China and other countries don't cheat to get ahead' [1].

The competing approach, largely reflected in current US policy, emphasizes domestic controls and competitive advantage. The Trump administration issued an executive order preempting state AI regulations in December 2025 [2] and moved toward export controls on model weights [3], which one observer characterized as making the US 'the world's most aggressive regulator of Expensive Maths' [3]. Analyst Zvi Mowshowitz argues the administration will not build a formal licensing regime, leaving governance dependent on opaque executive discretion—'winners chosen by the whims of those in power' [4]. Open-weight models at roughly 'Mythos-level' capability are now treated by both US and Chinese governments as posing credible national security risks, a threshold shaping both export controls and voluntary release decisions [4].

The Anthropic-Pentagon episode illustrates where these governance visions collide institutionally. Anthropic ended contract negotiations with the Department of War after the Pentagon insisted on blanket 'anything lawful' usage rights. Dario Amodei stated in February 2026 that this posture eliminates Anthropic's redlines on mass surveillance and autonomous weapons targeting, leaving 'no way forward' [4]. Separately, Anthropic and AE Studio developed GRAM (gradient-routed auxiliary module), a technique that routes dual-use capabilities into removable model compartments rather than suppressing them through safety training—a technical architecture that could allow selective capability unlocking depending on who controls the module [4].

Legal and governance analysts have examined whether labs can even coordinate on a pause. A Lawfare analysis argues a coordinated pause may be the most valuable safety intervention available, but antitrust constraints complicate direct lab-to-lab agreements without regulatory authorization [5][6]. The Centre for the Governance of AI has proposed an evaluation-based coordination scheme as a potentially workable legal path [7]. The Trump administration's national AI policy framework frames competitiveness rather than safety as the organizing principle [8][9], while Brookings has analyzed the broader structural competition between US and Chinese AI strategies [10].