Trump-Xi Meeting: Behind the Scenes AI Risk Talks

Why Informal Networks Are Driving US-China AI Diplomacy

Before the United States and the Soviet Union ever signed an arms control treaty, physicists and strategists from both countries were already meeting quietly in hotel conference rooms and university seminar halls — talking through scenarios no diplomat would yet touch officially. Those Track Two dialogues, running decades before formal agreements materialized, eventually shaped the Partial Nuclear Test Ban Treaty of 1963 and the architecture of strategic arms limitation that followed. Today, a strikingly similar pattern is emerging around artificial intelligence, as the Trump-Xi AI talks that may define the next decade of geopolitical competition take shape first in the shadows of academia rather than in any summit communiqué.

Ahead of a prospective meeting between President Donald Trump and Chinese President Xi Jinping, an informal web of dialogues has been quietly assembling at American think tanks and universities, drawing together researchers, former officials, and technical experts from both countries. The goal is neither a treaty nor a joint declaration. It is something more modest and, arguably, more durable: a shared vocabulary for discussing AI's most destabilizing risks and the beginning of a framework for managing them.

The precedent from nuclear history matters here. The Pugwash Conferences on Science and World Affairs, launched in 1957, gave scientists on opposing sides of the Cold War a neutral forum that governments officially denied while privately watching. Scholars such as Graham Allison at Harvard's Kennedy School have argued at length that such back-channel expert processes are often the only mechanism capable of building the epistemic trust required before formal diplomacy becomes possible. The same logic applies, perhaps even more urgently, to AI.

The Role of Think Tanks and Universities in AI Risk Talks

The current wave of US-China AI dialogue is being hosted not in foreign ministries but in institutions with enough academic distance to claim neutrality: research universities, nonpartisan policy centers, and science foundations. This is not accidental. Track Two diplomacy — the term coined by diplomat Joseph Montville in 1981 to describe unofficial, people-to-people engagement running alongside official state channels — thrives precisely where governments cannot afford to be seen compromising.

Think tanks such as the Center for a New American Security, the Carnegie Endowment for International Peace, and their Chinese counterparts at institutions like Peking University's Institute of International and Strategic Studies have become incubators for the concepts that eventually migrate into policy. Paul Scharre, Vice President and Director of Studies at CNAS and author of the widely cited work on autonomous weapons, has argued publicly that expert dialogue on AI risk is not a sign of weakness in competition with Beijing — it is a precondition for avoiding catastrophic miscalculation.

Universities add a different dimension. Technical researchers, rather than policy generalists, can engage Chinese counterparts on the actual mechanics of AI failure modes: misaligned objectives, opaque decision-making in deployed systems, the brittleness of large language models under adversarial conditions. That technical specificity is something formal diplomatic channels have struggled to accommodate.

Shared AI Risks That Both the US and China Acknowledge

Strip away the rhetoric of strategic competition and a narrow band of shared concern becomes visible. Both Washington and Beijing have reason to worry about AI systems that behave unexpectedly in high-stakes domains — financial markets, military command and control, critical infrastructure. Neither side benefits from an AI-triggered incident that escalates faster than human decision-makers can respond.

The scale of the competition makes the shared risk more acute, not less. According to the Stanford HAI AI Index 2024, China leads all other nations in AI patent filings, while the United States maintains a commanding lead in private AI investment, with the two countries together accounting for well over half of global AI research output. When two powers of that magnitude are deploying AI systems in proximity — whether in contested maritime corridors or in cyber operations — the probability of unintended interaction rises with each new deployment.

The specific risks that expert dialogues have reportedly brought to the table include: autonomous systems that misidentify targets, AI-generated disinformation that accelerates crisis escalation, and the potential for AI-enabled cyberattacks to strike infrastructure in ways that appear indistinguishable from acts of war. None of these risks belong exclusively to one side. An AI system that China deploys recklessly creates risks for China. The same is true in reverse.

What the Trump-Xi Meeting Could Mean for AI Policy

A formal summit between the two leaders would not, by itself, produce an AI governance agreement. The gap between informal expert dialogue and binding international commitments is vast, and the current political climate in both capitals makes ambitious multilateral frameworks difficult to negotiate. What a meeting could do is authorize the informal processes already underway — legitimize them, expand their scope, and signal that senior leadership in both governments considers AI risk a topic worth sustained engagement rather than a subject to be weaponized entirely for competitive posturing.

Marietje Schaake, International Policy Director at Stanford's Institute for Human-Centered Artificial Intelligence, has written and spoken at length about the importance of precisely this kind of signal. When heads of state publicly acknowledge shared AI risks, it creates political cover for the technical and policy experts doing the quieter work below. The Trump-Xi AI talks, even if their public output is modest — a joint statement, a commitment to maintain dialogue channels — could meaningfully accelerate what is already happening at the expert level.

Historical analogy is instructive again. The 1972 US-Soviet Incidents at Sea Agreement, which established protocols to reduce the risk of military confrontation at sea, emerged from years of quiet naval officer exchanges before it reached the treaty stage. It was narrow, technical, and almost invisible to the public. It also worked.

Obstacles Standing Between Dialogue and Real AI Governance

The obstacles are real and should not be minimized. Export controls on advanced semiconductors, restrictions on academic collaboration, and deep mutual suspicion about AI's military applications all constrain what informal dialogues can achieve. The United States has progressively tightened restrictions on technology transfers to Chinese entities, and Beijing has its own restrictions on data flows and research partnerships that limit what Chinese participants can share.

There is also the problem of verification. Any meaningful AI risk-reduction agreement would need some mechanism to confirm compliance, and the opacity of AI development — spread across thousands of companies, research labs, and government programs — makes verification far harder than counting nuclear warheads. Without verification, commitments risk becoming symbolic.

The asymmetry of AI governance philosophies adds a further layer of complexity. The United States has gravitated toward a mix of voluntary industry commitments and sector-specific regulation, while China has pursued more centralized algorithmic governance. These are not merely different approaches; they reflect genuinely different assumptions about the relationship between state authority and technological development.

What Comes Next for US-China AI Cooperation

The most realistic near-term outcome is not a treaty. It is the gradual institutionalization of expert dialogue channels — regular, semi-structured conversations between technical communities that build shared norms even without formal enforcement. That is how progress happened with biosafety protocols, with aviation standards, and with nuclear risk reduction hotlines during the Cold War.

The informal networks now meeting at think tanks and universities represent the earliest stage of that process. They are doing the definitional work: agreeing on what counts as an AI incident, what categories of AI deployment carry the highest escalation risk, what transparency measures might be mutually acceptable. That work is unglamorous and largely invisible. It is also probably the most consequential AI diplomacy happening anywhere right now.

Whether the broader Trump-Xi AI talks translate this groundwork into something durable depends heavily on political will in both capitals — and on whether both governments conclude that the costs of unmanaged AI risk outweigh the perceived advantages of unconstrained AI development. On the evidence of history, that calculation tends to shift when the risks become concrete enough. The question, as always, is whether the shift comes before or after the first serious incident.

Source: NPR Topics: News