Fault Lines: The Dual-Use AI Governance Vacuum in Asia

This is an ambitious and well-presented project addressing a real and important gap: the absence of a multilateral, binding, Global-South-inclusive architecture for dual-use AI governance.

The strongest contribution is the framework it proposes, which decomposes the governance problem into compute/precursor control, capability identification, use-case differentiation, national governance capacity, incident attribution, international verification, and consent/representation.

The live interactive tool is a strong presentation artifact and makes the analysis more usable for policy audiences.

The main weakness is that several load-bearing factual and causal claims seem overstated. The claim that the May 2025 rescission of the AI Diffusion Rule left “no replacement announced” and an ungoverned compute-access vector is too strong. The Malaysia/Huawei case does not truly support the paper’s thesis, at least as written. The paper should try to find a more nuanced description of the governance situation after the rescission. The claims regarding the Council of Europe Convention probably also need correction, at least greater detail/interpretation. Because the “codified asymmetry” claim relies heavily on this treaty, greater precision matters.

Relatedly, references to treaty negotiations should be narrowed and the WMD analogy is under-argued. The paper would be stronger if it specified which WMD governance functions transfer cleanly, which require adaptation, and which may not transfer at all.

The core gap identified by the paper is real and important. But the current version needs to reduce rhetorical certainty and support its causal claims more carefully.

The topic of the project is very timely. Things that could strengthen it include:

(1) Focusing on a narrower subset of the entire problem presented in the submission - currently, both the written document and the built online tool present too much high-level information (e.g. lots of references to policies, historical precedents, etc.) without depth of argumentation. It requires the reader to look for information instead of presenting a scoped out logic explicitly. Essentially, completing an 8-page paper and an online tool during such a short timeframe of a hackathon, as a single person team, seems to have taken away from the submission the ability to offer something deep enough to be concrete. Takeaway: focusing on one of the 4 contribution areas more in detail may have been more beneficial.

(2) The submission would benefit from a deeper exploration of the tensions between dual-use and diffusion, sovereignty, etc. for Asia. Currently, the submission is based on unexplained assumptions about their dynamics.

(3) Unclear audience and this takes away from the impact potential of the submission.

This paper analyzes seven dimensions of past WMD treaties to lay out a potential governance framework for AI. It concludes that current international frameworks have a carve out for AI in national security use cases and developing nations that rely on open-weight models lack governance over AI compute access. The paper also highlights a lack of incident reporting obligations, or audits of cross-border failures, or any mandatory pre-deployment safety evaluations in critical infrastructure. The accompanying website is a useful advocacy tool with great data visualizations and recommendations presented in a compelling manner.

I found the 4 cases meant to support the hypothesis that “Western export restrictions drove systemic integration of unmonitored AI into Asian critical infrastructure” ignore existing domestic and regional governance frameworks in Asia which weakens the argument that there is a “total vacuum” or no operational governance in place. The Malaysia example given seems to contradict the main argument where BIS guidance actually disincentivized the use of Chinese chips. It is not clear from the paper why the law would not apply to “Chinese-origin models” in the case of Singapore or Vietnam, and the argument could be stronger by referring to open-weight models more broadly.

The paper assumes that multi-lateral international treaty-based governance is the best path forward and lays out a potential framework. This works as an advocacy argument for a UN audience, but the limitations section fails to address the reality that negotiating an effective treaty will likely take over a decade, the end result may not benefit the global south, and that investment in domestic or regional solutions to compute governance in the meantime may have more impact. The paper’s claim that effective enforcement mechanisms are crucial can also be strengthened by addressing challenges with treaty ratification and universality given the region’s focus on sovereign AI.