Within the lead-up to Normal Secretary Xi Jinping’s September 2026 go to to the White Home to debate, amongst different matters, AI, U.S. Treasury Secretary Scott Bessent met with Chinese language Vice Premier He Lifeng at JPMorgan Chase’s New York Metropolis places of work. On the assembly, U.S. officers proposed establishing a dialogue between the USA and China, which might function a mechanism for notification of AI-related nationwide safety incidents. Bessent clarified the intent as: “We predict that identical to with any cross-border exercise, that shifting from opaque to extra transparency between the No. 1 and the No. 2 AI powers on the planet is essential.”
The Chinese language authorities readout of the following assembly between Xi and U.S. President Donald Trump acknowledged the 2 international locations “can proceed their dialogue on AI, alternate views on its dangers and advantages, and collectively stop the misuse and abuse of AI,” and affirmed that “AI should be stored below human management.” The latter assertion echoes the international locations’ 2024 settlement in regards to the function of human management in nuclear weapons, the place the USA and China agreed to “the necessity to preserve human management over the choice to make use of nuclear weapons.” In the meantime, the U.S. readout of the 2026 assembly famous that the USA and China agreed to ascertain the “U.S.-China Tremendous Intelligence (SI) Dialogue to alternate views on dangers and advantages associated to SI. The following alternate will happen by November 2026.” Additionally they agreed to type “a bilateral communication channel for SI incidents.”
What goal may this settlement serve? And what ought to the upcoming November 2026 alternate of views between the USA and China about AI deal with first (and what ought to it keep away from)? The proposed dialogues and communication channels wouldn’t restrain AI innovation on the a part of the USA or China, nor would they alter the underlying steadiness of AI energy between them. Additionally they don’t include any enforcement or verification mechanisms. In different phrases, they don’t characterize a type of AI arms management.
As a substitute, it’s extra helpful to judge these efforts via the lens of confidence-building measures: voluntary, non-binding measures geared toward decreasing opacity and facilitating communication between rivals in ways in which might, doubtlessly, scale back tensions and the dangers of inadvertent or unintended escalation — on this case, as they pertain to AI. The Chilly Battle historical past of confidence-building measures between the USA and the Soviet Union can shed some mild on the prospects of AI confidence-building measures between the USA and China within the modern period. However there are components of each the geopolitical context and the character of AI expertise that counsel new challenges and alternatives for confidence-building measures.
On the subject of the navy purposes of AI, the USA and China each understand a salient tradeoff. On the one hand, they’re locked in a broader strategic and geopolitical rivalry, of which fierce competitors — and even an “arms race” — over AI is an integral part. Certainly, as Trump opined within the context of U.S.-Chinese language relations, “whoever wins AI wins.” AI is a central part of China’s most up-to-date five-year plan, which calls on the nation to “seize the commanding heights of science and expertise growth.” AI gives doubtlessly transformative navy benefits, together with enhancing intelligence assortment and evaluation, decision-support, focusing on suggestions, command and management, autonomous maneuver and goal acquisition, and even enhancing routine “again workplace” features of navy organizations. That is solely underscored by current demonstrations of AI instruments on the battlefield within the conflicts within the Center East and Ukraine.
On the similar time, each international locations share mutual considerations about AI’s nationwide safety dangers, significantly the extent to which methods could also be implicated in choices about the usage of nuclear weapons (although China sees the safety dangers of AI from each a global safety and regime and social gathering safety perspective). As AI applied sciences are more and more included into the navy capabilities and decision-making means of two nuclear-armed nice powers, this creates new, doubtlessly destabilizing dangers. In a 2021 official place paper about regulating navy purposes of AI, for instance, China acknowledged the dangers that AI might pose for “the deepening of strategic miscalculation, undermining of strategic mutual belief, escalation of conflicts, and damaging international strategic steadiness and stability.”
Even when neither social gathering absolutely delegates management over the precise deployment of nuclear weapons to clever machines, the mixing of AI into the broader set of capabilities and choices associated to nuclear command and management, early warning, and launch choices creates new types of uncertainty and should exacerbate escalation dangers.
For instance, utilizing AI instruments to help decision-making round nuclear use or present varied programs of motion might bias choices in favor of extra escalatory choices. Analysis has proven that some massive language fashions (together with China’s DeepSeek mannequin) are extra inclined to suggest nuclear choices throughout simulated crises. Counting on AI for intelligence and early warning might lead to skewed (or, extra concerningly, hallucinated) intelligence evaluation, distorting assessments of impending nuclear threats. Simply this spring, a false, AI-generated intelligence report practically “began a warfare” between the USA and China in the course of the U.S. navy operation in opposition to Iran. Even when each the USA and China presumably need to “win” the AI race, neither would like to inadvertently or unintentionally stumble into an AI-triggered warfare (particularly a nuclear one).
One solution to handle these dangers is thru arms management agreements. Certainly, there are rising calls amongst some consultants and practitioners to work towards arms management for AI, particularly between the USA and China. However that is problematic, each due to international arms management traits and due to the character of AI expertise. With the expiration of New START in February and the emergence of a brand new, tripolar nuclear age, there seems to be little urge for food amongst nice powers for brand new arms management agreements. As a dual-use, general-purpose expertise, AI additionally presents distinct impediments to arms management, giving rise to each challenges in detecting dishonest and willingness to reveal capabilities. On prime of this, deep uncertainty in regards to the future navy implications (and benefits) of AI and the speedy tempo of innovation are prone to make nice powers reluctant to just accept any constraints on AI growth.
However arms management will not be the one possibility. Traditionally, confidence-building measures have emerged, even amongst rivals, in contexts the place states understand a shared curiosity in facilitating communication and selling transparency round navy capabilities to keep away from outcomes, corresponding to inadvertent battle, which are mutually understood to be detrimental. Confidence-building measures have been developed in the course of the Chilly Battle within the wake of the Cuban Missile Disaster for exactly this goal. In 1963, the USA and Soviet Union fashioned a bilateral hotline to institute a method of disaster communication. In the course of the 1967 and 1973 wars between Israel and its Arab neighbors, the USA used the hotline to speak with the Soviet Union to forestall miscalculations. As détente turned more and more fragile within the late Seventies, the 1963 hotline ultimately led to the 1987 settlement to ascertain Nuclear Threat Reductions Facilities throughout the U.S. State Division and Soviet Ministry of Protection. Extra lately, in 2013, the facilities’ function was prolonged to speak about crises associated to data and communications applied sciences, corresponding to cyber crises. Former U.S. President Barack Obama used the hotline on Oct. 31, 2016, to warn Russian President Vladimir Putin about cyber-enabled election interference.
One other notable Chilly Battle confidence-building measure settlement contains the 1972 Incidents at Sea Settlement between the USA and the Soviet Union to cut back the dangers of maritime crises. The 1975 Helsinki Ultimate Act, below the auspices of what was then the Convention on Safety and Co-operation in Europe, is one other paradigmatic instance. It has since been renegotiated and expanded a number of instances, most lately within the 2011 Vienna Doc.
Confidence-building measures aren’t a panacea for fostering stability between nice energy rivals, they usually typically lack the verification and enforcement mechanisms which are an indicator of arms management agreements. Their voluntary nature makes confidence-building measures comparatively simple to abrogate. As an illustration, following its full-scale invasion of Ukraine, in 2023 Russia introduced that it will cease sharing details about its armed forces per the provisions of the Vienna Doc. The method of negotiating confidence-building measures may additionally turn out to be an finish in itself, eclipsing a narrower (and doubtlessly extra fruitful) give attention to the tangible navy and safety advantages that confidence-building might accomplish. That stated, well-scoped efforts to facilitate communication between rivals, particularly throughout instances of disaster, typically characterize an enchancment from the established order. Furthermore, whereas confidence-building measures are unlikely to discourage aggression, observing deviations from shared expectations about conduct developed via confidence-building might present early warning.
Due to this fact, regardless of their constraints, the truth that the USA and Soviet Union might set up confidence-building measures in the course of the Chilly Battle to mitigate escalation dangers suggests some optimism a couple of U.S.-Chinese language settlement for AI. Nevertheless, 4 components distinctive to AI additionally counsel limitations round generalizing from the historic classes of Chilly Battle confidence-building measures to the current scenario.
The primary limitation is that, fortunately, a Cuban Missile Disaster-type of occasion has not but occurred for AI. Setting apart official skepticism in regards to the doom narrative that tech leaders, particularly in the USA, are promulgating, the truth is that there merely has not been a disaster between the USA and China triggered by AI that comes near the magnitude of 1962. The Cuban Missile Disaster was a formative second as a result of it crystallized a vivid, salient, shared sense of nuclear hazard, prompting efforts to institutionalize mechanisms to keep away from such risks even below heightened geopolitical tensions. Within the absence of an AI disaster of comparable magnitude, it’s not clear that each side agree on a shared set of nuclear dangers associated to AI, past the broad acknowledgment that sustaining some extent of human involvement in choices about nuclear use is essential.
Second, and relatedly, there may be little settlement between the USA and China about key phrases and ideas associated to AI escalation. For instance, each states have articulated totally different definitions of what may represent a deadly autonomous weapon (and, due to this fact, what class or kind of weapon may warrant higher restrictions and even an outright ban). China applies a excessive threshold, defining “unacceptable” deadly autonomous weapons as people who meet 5 standards, corresponding to appearing throughout the kill chain with an “absence of human intervention and management throughout your entire means of executing a process.” It isn’t clear that any present functionality satisfies all 5 standards. America defines such a system as one which, “as soon as activated, can choose and have interaction targets with out additional intervention by an operator.” There are additionally distinct ideas of the function of human beings in choices about the usage of power. U.S. coverage distinguishes between “acceptable ranges of human judgment” in choices about autonomous and semi-autonomous weapon methods (a extra permissive method), whereas calling for “human management” within the context of choices about utilizing nuclear weapons. In a 2025 white paper, China makes use of barely totally different language, affirming that “it’s important to make sure that human beings take the last word duty and that every one related weapons methods are below human management to forestall unauthorized actions.”
The third challenge is that, in contrast to nuclear weapons, AI will not be a weapon system. It’s a general-purpose expertise that not solely has manifold civilian purposes, however it additionally has diversified navy purposes related for nuclear and non-nuclear conflicts. Which means that conversations about the best way to assemble stabilizing measures to mitigate the potential escalation dangers stemming from AI, even when they’re narrowly scoped to give attention to the nexus of AI and nuclear weapons, should cowl numerous floor. Many AI purposes past the discrete concern about human management over the choice to launch nuclear weapons, as famous earlier, may additionally contribute to nuclear instability. On the similar time, many of those capabilities are additionally essential for the general navy competitors, limiting the prospects of voluntary restraints.
Lastly, the personal sector is an integral actor in any AI confidence-building measure effort. The prices, assets, institutional information, and ability essential to maintain continued innovation on the AI frontier imply AI energy is concentrated in a handful of companies that management entry to key AI inputs. Militaries are uniquely depending on these personal companies. But, these actors have their very own equities and pursuits in AI that won’t converge with authorities pursuits, particularly round potential confidence-building efforts that contain extra intrusive measures to examine AI methods on the design and growth stage or share delicate data.
In mild of those components, how ought to the upcoming bilateral conversations between the USA and China proceed? Policymakers ought to prioritize 4 particular traces of effort.
First, maybe counterintuitively, broader sensitivities within the U.S.-Chinese language relationship about AI might create house for each events to make progress round narrower, mutually shared considerations about nuclear escalation, offered that extra contentious points are understood to be off the desk. U.S. interlocutors, particularly, could also be tempted to broach matters that transcend the difficulty of AI-related nuclear dangers, strategic stability, and the best way to mature mechanisms to mitigate escalation. These might embrace different essential coverage challenges, corresponding to addressing disagreements about what constitutes AI security and safety dangers extra broadly, international governance of AI, export controls and restrictions on Chinese language AI fashions, and different associated areas the place there are important gaps between the U.S. and Chinese language positions.
The U.S.-Chinese language AI dialogue ought to give attention to constructing progress throughout the one clear space the place each events have already reached some type of settlement: human management over choices to make use of nuclear weapons. The problem will likely be the best way to assemble the dialog in such a means that avoids taking part in into Chinese language considerations about U.S. efforts to arrest China’s AI growth. Current calls from U.S. AI leaders, corresponding to Anthropic CEO Dario Amodei’s enchantment to “tempo the frontier,” aren’t conducive to progress on confidence-building as a result of they reinforce Chinese language perceptions that the USA is searching for to consolidate its benefits and prohibit extra open approaches or AI diffusion in ways in which may profit China. That could be the case (and will be the proper coverage), however it’s a distraction with regards to confidence-building measures to handle escalation.
Second, policymakers ought to purpose to construct on the prevailing basis of the 2024 settlement about human management over nuclear weapons, quite than increasing the scope of the dialogue. The easy assertion that people ought to management nuclear use choices leaves a lot unsettled and insufficiently outlined, significantly in mild of competing terminology and definitions each states have superior on this challenge. Due to this fact, the preliminary rounds of the dialogue ought to laser-focus on enhancing the readability and consistency of core ideas and definitions. Events ought to construct on the 2024 settlement to develop a confidence-building framework that extra clearly defines shared understandings about what constitutes “autonomous” capabilities and what human management means in apply, not just for the particular resolution to launch nuclear forces, but additionally for associated points round early warning and nuclear command, management, and communications.
Third, policymakers ought to house in on attending to widespread floor about observable conduct. The main focus must be much less on restrictions on the event of the expertise itself, however extra on working towards the event of observable indicators about the usage of AI in nuclear methods that each events agree can be harmful. Each side would profit from shared understandings in regards to the potential pathways to and off-ramps from escalation on the AI-nuclear intersection, and the observable patterns and indicators that the scenario could also be more and more on or off these escalation pathways — or, at a minimal, exchanging details about how all sides would understand various kinds of escalation situations. Such data could possibly be exchanged in a means that doesn’t reveal delicate details about AI methods, however is as a substitute centered on conduct and organizational and decision-making processes.
Lastly, how the U.S. authorities makes selections round regulating frontier AI labs and the companies that management associated AI infrastructure inputs generally is a type of confidence-building. Confidence-building measures don’t should be official agreements with rivals. Unilateral, voluntary declarations of restraint will also be a solution to construct confidence, and may run parallel to or help bilateral or multilateral agreements. Due to this fact, how the federal government opts to collaborate with (and management) the companies answerable for designing and growing AI methods related for navy operations, command and management, and decision-making is related for the prospects of confidence-building with rivals like China. Instantiating regulatory regimes or different frameworks for related personal actors round issues like incident reporting about accidents or unintended damaging outcomes, or sharing details about safeguards round system design for sure purposes, can buttress the credibility of adversary-facing confidence-building measures. In flip, failure to instantiate these domestic-level mechanisms might undermine the effectiveness of any future U.S.-Chinese language settlement.
Trying forward, there may additionally be alternatives to discover cross-domain confidence-building efforts in mild of the operational linkages between associated applied sciences and working environments, particularly AI, our on-line world, outer house, and nuclear methods. This implies different challenge areas the place the USA and China might understand widespread considerations, corresponding to across the cybersecurity of AI methods. Current examples in each international locations, corresponding to OpenAI brokers hacking Hugging Face or analysis demonstrating the usage of the AI instrument “WeWorm” to hack WeChat, spotlight these considerations. Profitable confidence-building measure efforts in a single challenge space could possibly be used as scaffolding to construct in different areas. However the focus within the quick time period must be on probably the most exigent dangers.
America and China are solely within the early phases of significant confidence-building measures to cut back AI escalation dangers. Furthermore, confidence-building measures aren’t a panacea for stabilizing nice energy relations (and have been by no means supposed to be). Nevertheless, the truth that each the USA and China already understand incentives to type confidence-building measures round AI and nuclear danger is a crucial place to begin.
Erica D. Lonergan is an assistant professor within the College of Worldwide and Public Affairs at Columbia College. This text was initially drafted for a Perry World Home convention on “Rising Expertise Agreements for Nuclear Threat Discount,” which was made attainable partially by a beneficiant grant from the Consortium to Scale back Nuclear Risks. The views expressed are solely the writer’s and don’t mirror these of Perry World Home, the College of Pennsylvania, or the Consortium to Scale back Nuclear Risks.
Picture: U.S. Division of the Treasury by way of Wikimedia Commons
