US Proposes China AI Safety Pact, Diverging from Arms Race

The U.S. proposal for a bilateral AI safety notification mechanism, presented by Treasury Secretary Scott Bessent ahead of the Trump-Xi summit, marks a significant shift in global technology governance. This move attempts to create a geopolitical "safety layer" atop the fiercely competitive AI development landscape, establishing a formal channel to de-escalate risks from unforeseen model breakthroughs. It strategically separates safety from the ongoing economic and military AI arms race, mirroring recent G7 discussions on AI principles but with the crucial inclusion of Beijing, seeking to prevent a catastrophic "race to the bottom" on safety protocols. The proposed notification system fundamentally alters the strategic calculus for leading AI developers like Google, OpenAI, and their Chinese counterparts such as Baidu and ZenseTime. It creates an implicit ceiling on reckless capability scaling, forcing a strategic recalculation where verifiable safety and containment become competitive assets, not just ethical afterthoughts. This framework advantages firms with mature, transparent safety governance while exposing the vulnerability of more opaque, state-influenced labs. The key "winner" is geopolitical stability, while the "loser" is the unchecked, growth-at-all-costs ethos that has dominated the last three years of AI development. Looking forward, the critical variable is enforcement and the precise definition of a "safety-relevant" breakthrough. Within six months, we can expect working groups to attempt to codify these thresholds, likely focusing on capabilities like advanced deception or autonomous replication. The real test will be whether this pact can survive the first major safety incident or a deliberate act of non-disclosure. This trajectory suggests the beginning of a two-tiered global AI ecosystem: a regulated core of state-aligned actors and an unregulated periphery, forcing a difficult choice for sovereign AI nations about which bloc to join.