AI Verification - A Role of Middle Powers

Hello, I'm Zac. I work at SASH, which is a Singapore-based field-building and research organisation. We work to test out new governance innovations with the Singapore ecosystem and broader AI safety world, and then try to export them globally with a focus on bridging Eastern and Western AI governance communities. The main thing that I'm working on at the moment is our AI verification team, and I'll be discussing why I think AI verification could be an important opportunity for middle powers. So what does the safety report say about AI verification?

Section 3.3 talks about hardware-based monitoring mechanisms. These are tools which rely on observing an AI system through reference to the chips that are running the system, or the data centres that are housing those chips. So instead of looking at the weights or the algorithms that the model is running on or evaluating its behaviour after it's been developed, you can look directly at the physical things that it's running on in order to find out things about the system. And in particular, verification is about demonstrating to another party certain facts about your system.

So this could be things like what data it was trained on, or what types of workloads are currently running or what models are currently running. I like to think about it like a metal detector for a data centre. So in the same sense that a metal detector tells you if there's a knife in your bag without making you empty everything that's in your bag and expose everything that's in there. A tool like a power meter installed outside of a data centre can tell you something about the workloads that are happening inside, without requiring you to expose the model weights or the user prompts or other sensitive IP.

I think verification could be very important for a couple of reasons. One, it can enable international agreements. So if the US and China want to agree on ways that AI should be developed or on guardrails around how it's deployed, these can enable those to be trustworthy. Two, it can enable safer AI adoption. If you're a company or a government that wants to adopt foreign-made AI systems, you might want high levels of assurance about what exactly you're getting. And three, it can enable countries without frontier AI capability greater sovereignty and insight into foreign-made AI.

Importantly, they don't exist yet and making them exist would be [inaudible] a huge win. Real quick, I want to talk about SASH's work. So we're a Singapore-based organisation and we work a lot in both the West and East, [inaudible] and what we're focusing on is building an international collaboration to develop these verification mechanisms. So if these mechanisms are to be trusted in international agreements, they need to be developed in multiple jurisdictions. Right now they're quite concentrated in especially the US and UK. So we are launching an international collaboration based out of Singapore to develop these mechanisms.

This involves building small-scale prototypes of what some of these mechanisms could be, and then demonstrating them to policymakers so they know sort of what verification mechanisms can do. And engaging researchers especially in Singapore and China. Come talk to me afterwards if you want to hear more about this. The main argument I want to make is that verification matters for middle powers. And there are three reasons for this. One, they can help you retain domestic regulatory capacity. So if you're going to adopt foreign-made AI, you shouldn't have to settle for whatever standards the company has opted to follow — EU or Californian standards.

These mechanisms can let you directly [inaudible] check whether a model has undergone certain evals, or whether it meets certain technical thresholds. Two. This enables you to more safely [inaudible] adopt foreign-made AI. So if you're an enterprise in a highly regulated industry or an industry where failure is catastrophic, you might want a greater guarantee about exactly what model you're being sold than just the label in the API endpoint, or just contract-based means. And the third thing is developing these mechanisms can help give middle powers a seat at the table, sort of in the same sense that the UK has been able to retain a strong seat at the table through leading in evaluation.

I think there's a big opportunity to do so in verification as well, and currently not that much effort is being put into developing these, and they can be done without kind of frontier-scaled resources. And also middle powers matter for making verification happen. So how is this? One is that middle powers can help support international agreements. If the US and China are to want to cooperate on how AI is trained or on guardrails around how it's deployed, it'd be very valuable to have a broad base of international experts who can help attest to how well these agreements are being followed.

If these mechanisms are only developed in the US or only in China, we have a lot less visibility into how well these agreements are being followed. And there's more political vulnerability. Two, middle powers can help directly develop these mechanisms. As I mentioned, there's a bunch of low-hanging fruit, and it can be done without sort of frontier-scale resources. I'll skip ahead on this because I'm a little low on time. But yeah, the main thing I'd like you to take away from this talk is that middle powers matter for AI verification.

And AI verification probably also matters for middle powers. And Australia might have a specific role. If you want to get involved: at 5 p.m., I'm hosting some networking drinks, sort of five minutes' walk away. So come along to that if you'd like to get involved more actively. I've linked some of our resources, some of our research and also some of our open roles here. Come find me.