US-China AI Race: Why Safety Talks May Already Be Failing

US and China Begin AI Safety Talks Amid Deep Distrust

The world's two dominant AI powers met this week for a structured dialogue explicitly aimed at preventing AI-related catastrophes from escalating into wider conflict. The US and China, whose competition now defines the contours of the US China AI race, sat down against a backdrop of escalating export restrictions, congressional hostility, and deep mutual suspicion. Before a scheduled Thursday-Friday meeting between President Donald Trump and President Xi Jinping, Treasury Secretary Scott Bessent confirmed that both sides had already begun preliminary discussions on establishing an AI safety notification mechanism.

The proposal, as Bessent outlined it, would create a dedicated channel through which either country could alert the other when its AI systems behave unpredictably or pose threats. The same channel, Bessent indicated, could be used to share each side's "vision of common goals." On paper, that sounds productive. In practice, analysts who study bilateral technology governance warn that the announcement's optimistic framing masks a fragile foundation.

The UN AI Advisory Body, which issued recommendations in 2024 calling for inclusive international AI governance structures, has long cautioned that bilateral arrangements between powerful states risk excluding smaller nations and creating frameworks that reflect great-power interests rather than global safety needs. The Global Partnership on AI (GPAI), a multilateral body with 29 member states, has similarly emphasized that durable AI safety architecture must extend beyond the two nations whose rivalry currently dominates the conversation.

Why China May Refuse to Trust US-Led AI Governance

China's silence immediately after Bessent's announcement spoke volumes. Where Washington was eager to publicize progress, Beijing offered nothing—a studied ambiguity that policy analysts read as skepticism rather than tacit agreement.

The mistrust has concrete policy roots. The Trump administration has steadily expanded export controls targeting China's access to advanced semiconductors, the physical infrastructure on which large-scale AI development depends. Applied through successive Commerce Department rulings and the Entity List, these restrictions have targeted the most powerful graphics processing units from major chip manufacturers. The stated aim: prevent China from training frontier AI models at scale.

Congress has compounded Beijing's wariness. Legislators have advanced bills that would further isolate China from global AI supply chains, treating technology transfer as a national security threat comparable to weapons proliferation. From Beijing's perspective, agreeing to an AI safety framework championed by Washington risks legitimizing governance structures designed, at least in part, to constrain China's technological rise. Participating in a "notification mechanism" proposed by a government simultaneously working to deny you the hardware needed for AI development is, at minimum, a complicated ask.

The 'AI Race' Framing and Its Strategic Dangers

The dominant metaphor shaping policy on both sides—that the US China AI race is a zero-sum competition with a single winner—may itself be the most dangerous variable in the equation.

Race metaphors carry embedded assumptions: speed matters more than safety, every advance by one side represents a loss for the other, and cooperation with a competitor signals weakness. Transplanted into AI development, these assumptions push both governments toward deployment timelines driven by strategic anxiety rather than readiness assessments. When a government believes it is losing a race, cutting safety evaluation corners becomes tempting. When both governments believe this simultaneously, risk compounds.

Researchers at institutions including the RAND Corporation and Georgetown University's Center for Security and Emerging Technology have documented how competitive framing in dual-use technology development historically accelerates timelines in ways that outpace safety infrastructure. The nuclear precedent is instructive. During the early Cold War, both the US and Soviet Union conducted atmospheric nuclear tests at a pace that prioritized strategic signaling over understanding long-term consequences—a pattern only interrupted after years of fallout data accumulated. AI systems capable of autonomous decision-making in high-stakes domains present analogous risks, potentially on faster timescales.

What Experts Say a Credible Global AI Framework Requires

A credible framework for managing the US China AI race cannot rest on a single bilateral notification channel announced at a trade summit. That assessment, implicit in the positions of multiple governance bodies, reflects hard-won lessons from arms control history.

When the US and Soviet Union established a direct communications hotline in 1963—following the Cuban Missile Crisis, which revealed how close the rivals had come to catastrophic miscalculation—the channel came with agreed protocols, verification mechanisms, and years of diplomatic groundwork. The Strategic Arms Limitation Talks of the early 1970s required not just political will but technical verification regimes, inspection rights, and defined terms for what constituted a violation. Without analogous specificity, an "AI safety notification" mechanism risks becoming a diplomatic talking point rather than an operational safeguard.

Governance researchers consistently identify three minimum requirements for effective bilateral AI safety infrastructure: shared definitions of what constitutes a dangerous AI incident; independent verification that either side has flagged relevant events; and agreed escalation procedures if a notification triggers dispute. None of these elements were present in Bessent's announcement. The UN AI Advisory Body's 2024 report explicitly warned against frameworks that lack accountability mechanisms, a concern that applies directly here.

Comparing the US-China AI Rivalry to Past Tech Cold Wars

History offers some comfort—and a great deal of caution—for those hoping the US China AI race ends more cooperatively than it began.

The original Cold War featured decades of deep ideological hostility, yet the US and Soviet Union eventually negotiated the most ambitious arms control architecture in history. The 1972 Anti-Ballistic Missile Treaty, followed by SALT I, SALT II, and the Intermediate-Range Nuclear Forces Treaty, all emerged from a relationship characterized by profound mutual distrust. The common thread was not goodwill. It was mutual vulnerability. Both sides recognized that unconstrained competition increased the probability of outcomes neither wanted.

AI presents a structurally different problem. Nuclear weapons are material objects—countable, photographable by satellites, constrained by physical production capacity. AI capabilities are software, trained on data, deployable globally through networks, and extremely difficult to verify or limit through traditional arms control tools. The export controls Washington has imposed on high-end semiconductors represent an attempt to reimpose physicality—constraining AI development by controlling the hardware it requires. Whether that approach succeeds or simply accelerates China's domestic chip industry, as some analysts predict, remains one of the central unresolved questions of the rivalry.

The Stakes: Why AI Governance Failure Endangers Everyone

If the current round of talks collapses without producing durable mechanisms, the consequences extend far beyond the two capitals conducting them.

Advanced AI systems deployed in critical infrastructure, financial markets, autonomous military platforms, and public health surveillance can fail in ways that cross borders instantly. A misconfigured AI system managing power grid operations, a model disseminating false emergency alerts at scale, or an autonomous system misidentifying a target—none of these scenarios respect bilateral diplomatic frameworks. All become more likely in an environment where the US China AI race rewards speed over safety.

GPAI has repeatedly warned that a governance vacuum at the top of the AI development hierarchy will be filled by fragmented national standards, making eventual interoperability and accountability harder to construct. Smaller nations, lacking resources for frontier AI safety research, are particularly exposed. They will inherit whatever safety culture the major powers establish—or fail to establish.

The talks this week are not nothing. Any open channel between Washington and Beijing on AI is preferable to silence. But mistaking preliminary contact for a functioning safety framework would be the most dangerous misreading of the moment. The US China AI race is real, the structural risks are documented, and a notification channel announced at a trade summit is not, by itself, governance.

Source: Ars Technica - All content