Research Post

The Case for Including Other States in US–China AI Verification

The participation of other states can strengthen a bilateral AI verification regime between the US and China by augmenting implementation capacity, expanding monitoring coverage, and lending broader international legitimacy.
JSJeba Sania SASumaya Adan JWJoanna Wiaterek November 2026

In September, following high-level talks on managing frontier AI safety risks, the United States and China agreed to open a bilateral channel for AI-related incidents. The agreement contains no provisions for verification. Tensions were present even before the talks. Anthropic CEO Dario Amodei published an essay on the global risks of Chinese AI, and China’s state-run Global Times dismissed it as an attempt to hinder Chinese innovation. This lack of trust is a fundamental barrier to cooperation, and it stems in part from an inability to substantiate one another’s claims about AI development.

This is the problem “AI verification” is meant to solve: the ability of developers, third parties, or states to verify claims about an AI system’s use, development, behavior, and capabilities. Such ability can improve trust between actors and demonstrate compliance with treaties and legal obligations. This piece addresses international verification, meaning how states could verify compliance with an international agreement restricting AI development and deployment. Within that, verification mechanisms are specific approaches for assessing compliance. A verification regime is the broader institutional arrangement, including the agreement, the mechanisms, and the actors that operate and judge them.

As concerns over the risks of frontier AI development grow, international verification has emerged as a research agenda that examines whether states have the technical capacity and political will to negotiate and implement enforceable agreements. Without verification, states may hesitate to enter into AI agreements in which compliance is difficult to assess, compromising sustained international cooperation.

In this piece, we argue that strategically including additional states can strengthen, rather than dilute, a US–China-centered regime. Broader participation can make a regime more robust by plugging implementation and coverage gaps and adding legitimacy to a bilateral agreement. Concretely, the involvement of states beyond the US and China could:

  1. Provide additional implementation capacity, reducing the enforcement costs for the US and China;

  2. Expand monitoring coverage, reducing opportunities for circumvention and jurisdictional arbitrage (taking advantage of differences between legal regimes);

  3. Add legitimacy, which bolsters the durability of a regime.

Throughout, we surface key gaps in existing proposals, examine how other states could help address them, and draw on current and historical analogues as practical examples.

Existing proposals tend to focus on verification mechanisms, meaning what to measure and how to detect evasion, rather than on regime design. They generally envisage combinations of technical monitoring, inspections, hardware controls, national intelligence and, in some cases, dedicated international institutions. Many are designed around two rival states attempting to verify each other’s compliance in a privacy-preserving manner, with one party attempting to evade detection as the threat model. These frameworks focus on one party monitoring the other, not on how monitoring capacity could be shared across more states. The question of political legitimacy, meaning which entity has standing to judge compliance, remains underexplored, as many proposals assume the existence of a regulatory body. Implementation is attributed either to entities with technical capability, such as compute providers or frontier AI labs, or to those without a conflict of interest, such as domestic regulators or an international governing body that does not yet exist.

While some frameworks are formally country-agnostic, the United States and China frequently serve as the central strategic case. Scher and Thiergart, for example, explicitly frame verification around how the US and China might credibly verify one another’s compliance with a bilateral agreement, while Baker et al. use a hypothetical US–China agreement to illustrate a layered verification architecture combining technical monitoring with national intelligence. More institutionalized proposals go further: Scher et al.’s proposed agreement on advanced AI places the US and China at the center of the regime’s governance and independent verification functions.

Some justifications for a US–China-centered verification regime are summarized below.

  1. Compute concentration: Global compute is concentrated in the US and China, the main actors developing frontier AI systems. Compute refers to the software and hardware infrastructure used to train, run, and deploy AI systems.

  2. Tractability: International agreements are easier to negotiate when fewer countries are involved, as less effort is needed to build political will and implement the agreed provisions.

  3. Lack of time: Concern that the window to implement sufficient safeguards may be short has prompted proposals to prioritize near-term, workable agreements focused on the states most consequential to frontier AI development.

  4. Influence: Other states will enter verification regimes once the US and China have done so, given their influence on the international system.

Across this literature, the US–China relationship is treated variously as the motivation, threat model, and institutional core of verification, while the roles of other states remain comparatively underdeveloped. However, conventional understandings of how international agreements are negotiated, ratified, and implemented often rest on assumptions inherited from the unipolar era that may no longer hold.

Our argument supports strategic broader participation, not membership for every state, and broader participation need not mean slower talks. The US and China would still negotiate the agreement’s core terms between themselves. A small group of states would then be invited for specific functions, such as testing hardware or hosting monitoring stations, with no veto over those terms. Sensitive information would be shared on a need-to-know basis, under confidentiality rules similar to those the IAEA applies to safeguards data. Negotiations stay bilateral, while the capacity and credibility behind them widen. As this analysis is exploratory, we do not provide a comprehensive framework for an international regime. Rather, we aim to spark deeper engagement on this topic.

Neither superpower has the full capacity to self-certify or enforce a regime

Jurisdictions with expertise in semiconductor fabrication, metrology, or industrial certification, such as the Netherlands, Taiwan, Japan, and Germany, could supply critical testing capacity. Ongoing verification research points to specialized technical capabilities such as certifying hardware attestation, auditing firmware, and analyzing chip provenance. Much of the expertise and infrastructure relevant to these functions sits in semiconductor supply chains outside the US and China. Consequently, neither is positioned to self-certify credibly from a technical and capacity standpoint. A useful analogy is the OPCW’s network of designated laboratories across multiple member states. Rather than requiring parties to validate their own compliance, this approach distributes analysis. Because most leading testing jurisdictions are aligned with the US, an AI regime could likewise split testing between laboratories that each side accepts.

Broader participation could also raise the costs of non-compliance by enabling coordinated responses across multiple jurisdictions. Such coordination provides significant leverage against states that are economically interconnected within the international system. In 2010, for example, the UN Security Council imposed sanctions on Iran over its noncompliance in pursuing its nuclear program. These sanctions, reinforced by additional US and EU measures, are widely understood to have led Iran to negotiate the Joint Comprehensive Plan of Action (JCPOA).

Other states can broaden monitoring coverage to reduce circumvention, gaming, and jurisdictional arbitrage

Weak spots outside a regime’s intended perimeter are the default channels for circumvention. Data-center capacity is distributed across jurisdictions beyond the United States and China. Failure modes include other compute-hosting states serving as evasion routes and smuggling networks. During the Cold War, for example, Coordinating Committee for Multilateral Export Controls (CoCom) restrictions on the Soviet Union pushed dual-use technology trade toward neutral third parties such as Switzerland and Sweden. Proposed accountability mechanisms, such as whistleblower programs and licensing, face the same jurisdictional limits.

A regime whose oversight depends on actors remaining within monitored jurisdictions may also unintentionally create incentives for smuggling, already the subject of US prosecutions, or for actors to game the regime. Offshore banking and profit shifting to tax havens offer analogues. Frontier firms may move their operations, or parts of them, to less-regulated regions to exploit regulatory gaps.

To mitigate these risks, the regime could adopt a scaled-down version of the Comprehensive Nuclear-Test-Ban Treaty’s (CTBT) International Monitoring System, in which sensor data from verification stations in dozens of third-party states is transmitted to an international body rather than controlled exclusively by the states being monitored. While international bodies take time and effort to stand up, a small body of the most consequential countries in the frontier AI supply chain might be more suitable given current time pressure. Other jurisdictions, such as South Korea, which is critical to the frontier AI supply chain as a manufacturer of advanced memory chips, may be particularly helpful in providing early warnings about illicit smuggling networks. In 2024–2025, for example, authorities in Singapore, a semiconductor and logistics hub, helped uncover smuggled Nvidia chips destined for China.

Multinational efforts are harder to dismiss as politics

States with working ties to both sides, such as Malaysia, can join multinational inspection bodies that prevent the US or China from dominating an inspection team and guard against accusations of political interference. Such bodies can consist of technical experts who are civil servants with an independent mandate to preserve political neutrality. This can prevent verification findings issued unilaterally by either the US or China from being perceived as politically motivated. Such misperceptions risk escalating tensions and eroding trust, which could ultimately break down the regime altogether. The OPCW’s 2013–2014 mission to verify and destroy Syria’s chemical weapons stockpile shows how multinational efforts can distance verification from the political interests of any single state. Other states could also serve on a verification oversight body, similar to the IAEA Board of Governors, judging whether the facts show compliance rather than setting the agreement’s terms.

If bilateral tensions escalate and credibility declines, the inspected party can withhold site access, dispute findings, or decline to act on inspection mandates. Other states then have a critical role in mediating disputes and rebuilding trust to uphold the bilateral agreement.

It may also become politically desirable to host data centers in third-party states, creating neutral sites for independent monitoring. Data centers could be hosted in states that maintain working relations with both the US and China, such as Saudi Arabia and Singapore. Because neither the US nor China could control the facility or manipulate its findings unilaterally, both may view the resulting assessments as more credible. Such an arrangement could help a regime withstand a deterioration in US–China relations: because inspections occur outside both countries’ jurisdiction, neither government could unilaterally revoke the other’s access.

Conclusion

Current verification proposals largely leave other states out, yet their participation can close the implementation, monitoring, and legitimacy gaps that a purely bilateral regime leaves open. Managing AI risks will require the US and China, but waiting for the two to set aside their rivalry is a gamble with time we may not have. Further work should set participation criteria and examine the costs of broader participation, the effects on non-parties, and capacity gaps among participating states. Broader participation will not end great-power competition, but it can keep international AI governance from becoming hostage to it.

Cite as: Sania, J., et al. "The Case for Including Other States in US–China AI Verification." Centre for AI Security and Access, Nov 2026. https://www.casa-ai.org/the-case-for-including-other-states-in-uschina-ai-verification

The views expressed are those of the author(s) and do not necessarily reflect those of CASA or its funders.

← Back to all Blog posts