Skip to content
VECTOR WIREAI INTELLIGENCE
UTC
Home

US-China bilateral handshake on military AI fills a vacuum Geneva left empty

The week's two most consequential developments in military AI governance moved in opposite directions. In Geneva, more than 120 states at the UN Review Conference on Conventional Weapons failed to agree on a mandate for a legally binding treaty on autonomous weapons, collapsing just before 5 a.m. on Saturday after all-night talks could not overcome Russian objections1,3,5. In Shenzhen, days earlier, Donald Trump and Xi Jinping agreed that humans should always be involved in military decisions and that AI should not be allowed to act autonomously on major military actions, according to Treasury Secretary Scott Bessent6,8. The bilateral handshake between the world's two largest military AI investors is now the closest thing to a shared statement of intent on autonomous weapons — not because it carries normative weight on its own terms, but because the multilateral process that was supposed to produce binding rules did not. Whether the Shenzhen exchange amounts to more than rhetoric depends on definitions neither side has yet published.

Vector Wire reported on November 22 that the Geneva talks yielded only a non-binding mandate[6], and on November 21 that the Trump-Xi agreement had been reached[7]. The argument here advances both: the relationship between the two outcomes is what matters, because it determines who writes the rules.

The Geneva conference asked countries to "elaborate by consensus" a possible instrument on lethal autonomous weapons systems, a formulation that may not carry any legal obligations. The compromise gave states two years to propose regulations, with negotiations on adoption and enforcement likely to drag on well beyond that, and room was left for some of those yet-to-be-adopted regulations to remain optional2. The ICRC had called for legally binding rules within a year, a timeline backed by UN Secretary-General António Guterres and Pope Leo. What emerged was, by any measure, far weaker.

Russia led the obstruction. It deployed what the Guardian described as "uniquely obstructive stalling tactics from the opening day". But Russia was not alone in resisting binding commitments. The United States, too, was reluctant to commit to a legally binding protocol protecting civilians from autonomous weapons. Both countries have invested heavily in AI warfare. Both Russia and Ukraine are using AI to track targets and plan attacks in the war now entering its fifth year. The U.S. has made a major push to increase its use of AI in combat and is testing advanced fighter drones that fly themselves.

ANALYSIS Washington's posture in Geneva and its posture in Shenzhen are consistent, though not necessarily coordinated. The United States has resisted legally binding autonomous-weapons rules across multiple administrations; the Geneva obstruction is not new. What is new is the bilateral channel with Beijing. Blocking a multilateral treaty preserves freedom of action. A bilateral understanding framed in broad language about human involvement in "major military actions" costs little in operational terms while signaling restraint to domestic and allied audiences. The Shenzhen exchange did not cause the Geneva collapse, but it offers Washington a rhetorical alternative now that the collapse has occurred.

Bessent's characterization of the Shenzhen exchange deserves scrutiny on two levels: reliability and substance. There is no joint statement, no signed document, and no published text of the agreement — only Bessent's paraphrase of a verbal exchange on CNBC. Bessent told CNBC on November 20 that former Nvidia CEO Jensen Huang had not come up in Trump's talks with Xi7. Hours later, Trump himself said he had discussed Huang with Xi, and that Xi called Huang "a brilliant man". The Treasury Department did not immediately respond to CNBC's request for comment on the discrepancy. If Bessent misreported whether chip-related discussions occurred in the same meeting, the military-AI paraphrase he offered from that meeting cannot be treated as a precise record of what was agreed. What he described is at most a verbal understanding whose precise terms are unknown.

Even taken at face value, the language does less than it appears to. Bessent told CNBC that Trump and Xi "agreed that humans should always be involved in any military decisions that SI makes, and the SI should not be allowed to be autonomous for major military actions". The qualifier "major" does real work: it leaves undefined what constitutes a minor military action, and whether autonomous systems operating below that threshold require human oversight at all. Bessent used the administration's preferred term "SI," for super intelligence, rather than AI, a framing that narrows the scope further by implying the guardrails apply to a class of technology that does not yet exist in deployed military systems. An understanding about superintelligence cannot constitute an operational norm for the autonomous weapons already on battlefields today; at best, it is a placeholder for a future conversation.

Bessent also stressed that the United States "must stay ahead" of China in AI, and that "our best protection against the foreign SI and exploits by foreign actors is for our labs to maintain their substantial lead". The administration's theory of military AI safety, as articulated by Bessent, rests on American technological dominance rather than on negotiated constraints, a position structurally incompatible with the binding multilateral framework Geneva was trying to build, and consistent with the dominance-first posture Vector Wire has tracked since Bessent's bid to become the administration's AI czar[5].

Nvidia, under interim CEO Colette Kress following Jensen Huang's ouster, had days earlier announced it would no longer ship H200 chips to China[1]. Bessent suggested that U.S. policy allowing the sale of Nvidia's H200 AI chips to China remains unchanged, a statement that sits uneasily alongside Nvidia's own decision to exit those sales under its new board chair Harvey C. Jones.

Dr. Cordula Droege, the ICRC's director of international law, movement and partnerships, warned that the window for regulation is closing. "We're starting to see artificial intelligence … being fielded in the battlefield and perhaps also already in weapon systems. If we leave this unregulated, it will be much more difficult to put the genie back in the bottle," she said. Max Tegmark of the Massachusetts Institute of Technology argued that states have agreed it is sufficient to just talk about the issue for the 14th year while these weapons enable atrocities.

Tegmark's frustration points to the structural problem. The need for consensus at the Convention on Certain Conventional Weapons forced the conference to adopt a much weaker option despite an overwhelming majority of states wanting a legally binding treaty. Russia demonstrated the power of that consensus requirement in Geneva. The bilateral channel between Washington and Beijing faces no such procedural constraint, but it also carries no enforcement mechanism, no verification regime, and no obligation to third parties. An understanding with no definition of its own scope, no enforcement, and no third-party obligation is, on its face, indistinguishable from a talking point. More than 100 states had been pushing for a new law, saying autonomous weapons raised entirely new challenges. Those states are now spectators to a verbal exchange between two powers whose key terms remain unspecified.

European and NATO states, including Ukraine, as well as China and Iran, had expressed support or willingness for a new protocol at the Geneva conference. China's dual engagement complicates the thesis that Shenzhen fills a vacuum Geneva left. Beijing has not abandoned the multilateral track; it endorsed a new protocol even as it struck the bilateral understanding with Washington. The two channels coexist rather than substitute for each other. But coexistence gives Beijing optionality: it can point to the Shenzhen exchange as evidence of responsible behavior while the Geneva process stalls, without accepting binding constraints from either venue. The vacuum is not that multilateral diplomacy has ended; it is that neither track currently imposes enforceable obligations on the states building the weapons.

The Geneva outcome set a two-year window for countries to propose regulations on lethal autonomous weapons systems. Whether that timeline produces anything enforceable depends on whether the consensus requirement can be navigated past Russian objections, and whether the United States, having secured its bilateral understanding with China, has any incentive to accept multilateral constraints it blocked in Geneva. Bessent's framing of U.S. lab dominance as the country's "shield" against foreign AI threats suggests the administration's answer, for now, is no.

The two-year elaboration window is the next concrete marker. Three things will determine whether the Shenzhen exchange amounts to anything more than rhetoric: whether Washington and Beijing publish operational definitions of "major military actions" and "human involvement"; whether either side alters procurement or deployment doctrine in observable ways; and whether the bilateral channel produces any mechanism for verification or third-party transparency. Absent all three, the exchange is a talking point dressed as a norm, and the regulatory vacuum persists regardless of which track claims to be filling it. The ICRC's Droege put the stakes plainly: the technology is already on the battlefield, and the regulatory architecture is not.

Was this useful?

The day’s vector, in your inbox.

Vector Wire Daily: one dispatch each publishing day.