An argument for accelerating international AI governance research (part 1)

After a good long while of reading, thinking, and working at the edges of AI governance (I am an historian of international law and global governance as well as an AI capabilities evaluator and qualitative red-teamer), I am finally taking the plunge and starting to write about it. 

My main contention is this: It seems to me that now, more than ever, we need to rapidly expand the field of AI governance and, specifically, work on its international dimensions. Right now, world governments and international institutions, as well as AI labs developing AI systems that will (and are) finding private and public applications across the world, are in dire need of this talent. Yet the pipelines for producing the talent necessary to implement international AI governance are not sufficiently developed to meet that demand. Without that talent, there will be measurably little progress in efforts to achieve any international AI governance, and what success we do have may be misdirected and therefore ineffective.

If we cannot solve for today's AIs, which are orders of magnitude less complex than advanced AIs that seem in-range by 2030, we may be highly unlikely to be able to implement any meaningful international governance mechanisms that we do happen to theorize later on.

We must start taking this international governance gap much more seriously, and now. 

In this and a series of subsequent posts, I hope to lay out this case as comprehensively as I can, albeit from my own limited perspective. The trajectory of my argument, for now, is the following:

1. International governance is a high-viscosity fluid
2. International governance is responsive to certain authorities
3. International governance requires institutional coordination
4. International AI governance can help reduce suffering risks today
5. International governance may have to change considerably to accommodate advanced AI
6. A reassessment of prevailing international governance models as they pertain to AI and AGI
   1. On “strong” versus “weak” internationalisms
   2. Reconsidering the nonproliferation analogy
   3. The role of INGOs
   4. Where is the "compliance surface" of international AI governance?

I want this to be as useful to others as I can make it, of course. If there is a topic you'd like to see sooner than others, please comment and/or vote below. 

International governance is a high-viscosity fluid

I want to begin my argument with a simple point. 

We have to appreciate in greater depth a fundamental property of international governance: it is highly viscous. It moves slowly. And it requires a great deal of sustained effort to generate sufficient momentum.

This viscosity arises from three sources:

1. Even today, in the afterglow of the post-war human rights revolution,  international governance inherently proceeds on a voluntary basis. Roughly speaking, any such governance depends on mutual agreements between equals (legally speaking, at least). And while sometimes those agreements are subject to oversight and arbitration from international entities (like UN bodies) or other states and INGOs, there is no authority that can absolutely compel a state to honor international agreements. This, still, is the domain of natural law. 
2. International agreements are often formalized by treaties or conventions. However, these treaties or conventions are often subject to ratification and implementation by the municipal authorities of signatory states. In the United States, the Senate must ratify any treaty negotiated by the Executive Branch. And, despite current scholarship on the 'secret Congress' that reassures us that important business still manages to get done in hyperpolarized times, this is no guarantee that international agreements will stick. For instance, the U.S. Senate infamously voted not to ratify the Treaty of Versailles (in which allied states normalized relations with Germany and formed the League of Nations), which prevented the United States from joining the League. The U.S. and Germany only made peace in 1921.
3. International governance initiatives, whether achieved through institutions, conventions, or treaties, compete for political will and attention with several other high-profile needs. In other words, it takes substantial amounts of time, resources, and willpower to reach even a singly multilateral treaty agreement. A pertinent example is the Iran nuclear treaty formalized in July of 2015, marking the end of a nearly 12 year combined effort to achieve the P5+1 framework. I was at the Wilson Center in Washington, D.C. when the success of the treaty was announced, attending a workshop. I asked a leader of the workshop, also a longtime mentor,  what he thought of the news, and he replied, with tears forming in his eyes, "I never thought I would live to see it."

Time is not on our side. We must start building international frameworks, institutions, conventions, and will. That, in my opinion, begins by expanding the AI governance talent pipeline.

There is difficult work to do. There are exceedingly few solutions to problems of AI governance, and the window for effecting them may be closing faster than we realize. This is especially true for projects of international governance, which regard only certain sources as authoritative. 

In my next post, I will return to explore these claims in more detail.

(image: Carel Willink, “Late visitors to Pompeii,” 1931)