Ben Bucknall – Open Problems in Technical AI Governance

Transcript

 Nice. Thanks very much. I'm Ben, currently based at the Centre for the Governance of AI, and I'm going to be giving a rough overview of a paper, which very excitingly is coming out next week, hopefully up on arXiv, on Open Problems in Technical AI Governance. And this is work done in collaboration with Anka Reuel at Stanford and a bunch of other people, some of whom are in this room. So thank you very much to them.

What is technical AI governance? So there's a bunch of different definitions that you may want to use here. We settled on this one, which is “technical analysis and tools for supporting the effective governance of AI.” This includes a bunch of different activities that people might be carrying out; for example, looking into the feasibility of different proposed policy actions, or as Helen talked about earlier, filling in the details of proposed policies, or in cases where we don't have sufficient technical solutions or implementations for policies, working on technical methods and tools that can advance our ability to implement them.

Formalizing that a bit more, we lay out in the paper three key ways, which we term Identify, Inform and Enhance. Identify is stuff along the lines of better visibility into AI and its development that may motivate the need for governance intervention. Informing key governance decisions, this is stuff where policy makers may have a bunch of different options to choose from, and the efficacy of those different options may depend on technical details.  And finally, Enhance, which is what I spoke about earlier, where we may need further technical tools to implement and enforce policies. 

I think one thing just to note here on this slide is that, we're not advocating for a position of techno-solutionism here, and we view technical AI governance as merely one component of a comprehensive AI governance portfolio. We view it strictly in service of socio-technical, political and other solutions.

So in the paper, we taxonomize technical AI governance according to what we term targets and capacities. Targets are key elements of the AI value chain, going through from data and compute to a system's deployment in the real world.  Capacities refer to activities that can enable stakeholders to understand and shape the development, deployment, and use of AI systems. The first four of the capacities can be applied to each of the four targets, and the last two - operationalization and ecosystem monitoring - are more general cross-cutting capacities.

So just to give a few examples of the open problems that fit in the different categories here. We've heard quite a bit about evaluations today, so for example model evaluations would be in models intersecting with assessment. We also have other stuff. So, verification and data; this could be stuff like being able to verify that a given model was trained on a given data set. In verification and deployment, we have stuff like watermarking of outputs or detecting AI-generated content.

Just got a bunch more example open problems here, but I see that I'm short on time, so I won't go into more detail. But what I will do  is provide you with a little sneak peek of the paper. If you want to follow the QR code -  it's not yet on arXiv - we're waiting on a manual review of it. But if you follow the QR code, then you can have a look for yourself a week ahead of time. I'll leave it there. There's 10 seconds left for you to get that up on your phone.  Thank you.