Opening Keynote on AI Policy & Governance

Kimberlee Weatherall

Thank you very much. It's a pleasure to be here. Liming and Tiberio have painted us a picture of expanding, accelerating, super accelerating AI capabilities and deep safety challenges. And the word I keep hearing is speed, speed, speed, speed. This is very uncomfortable for lawyers. So, you know, I’d just like to acknowledge this moment as a governance person, that this is a very uncomfortable place. Look, with this, event being built around the international AI Safety Report, I should note that, you know, in terms of anchoring, you might notice I haven't highlighted any sections of the report.

And that's because policy and governance is nowhere and everywhere in this report. And the report is designed to inform policymakers. It is a scientific review and an attempt to build some consensus on common ground around what the science is. It is not an attempt. It is not a scientific or consensus review of the laws, policies or governance approaches that we need to address these. It's kind of like, you know, the big sort of scientist hospital pass. Here's all the problems: go write a law. But of course, we couldn't actually do some kind of consensus approach in law, policy and governance.

But in reality, consensus occupies an interesting place in law and policy. If you ask for consensus, you're actually giving everyone veto power. And so what? In reality, we don't always achieve it. We don't always aim for it. People don't always have to agree when it comes to law and governance and policy in AI. That's kind of the point of being people and being people in a democracy and being people in a global world where people and nations come from different political and cultural traditions. And that's why with this week's event in Geneva, that's already been referred to as the UN event.

It isn't a treaty making, and it isn't a consensus. There's no appetite for a new treaty or a new global body. Instead, the way that they framed it as a global AI governance dialogue. And I actually think that's a really important idea for us to hold on to. What we need is ongoing dialogue. That's actually how law and policy can respond at the kind of speed that is being demanded: by constant engagement. And what we can and fervently hope for is an effort to find, not consensus, but areas of common ground.

And more importantly, given the speed, ways that we can cooperate to address the kinds of things that are coming down the line (and rapidly). And the key thing I'd like you to know, actually, is that a lot has been done. You know, I could stand here and tell you we don't have a new AI law. We don't have it in law in Australia. We don't have an AI law in many countries. But actually quite a lot of progress has been made on trying to find both the basis for a common ground around AI governance

or moments of common ground. And the places where these things and these discussions can happen. So, you know, over there, we have Bletchley Park summit of 2023 and its successors. We have the international AI Safety Report itself. And of course, coming out of that, also that network of AI safety institutes, doing a lot of hard work to give us a sense of where the science is. On this other side we have all these other initiatives that have been going on. We have work on transparency mechanisms, work

on technical standards, and various other kinds of interventions, including, importantly, interventions, not just from the technical world. Right, we have interventions from the UN's cultural, social and economic body, UNESCO. We have an intervention, of course, recently from the Vatican talking to us about AI, including AI safety. And, while there's not perfect alignment between the kinds of principles and ideas that are being discussed in here, there is, in fact, more common ground than you might think around the need for risk management, the need for ongoing and lifecycle monitoring, human-centred approaches.

And then of course, this week we do have the UN stepping in with this first global dialogue on artificial intelligence governance. And that's important because in that dialogue that's the UN running it. It means that every country is there. Now, if we were to put all these developments on a kind of timeline, it would be messy. It would be messy. But what we might see is a pattern from the early ethics principles through Bletchley and discussions led by a small group of advanced economies expanding out further into what we could truly now say is a genuine global conversation.

And that's also important. We are expanding out the conversation, and we're expanding that out to different groups. All these people who were affected and not just not over a period of decades, over a period of, you know, 3 to 4 years. That's lightning speed in law terms, by the way. So, you know, I think we should recognise that this is an achievement that these conversations are happening, that the places where people can talk and respond and the connections between people to talk and respond to are being formed.

It doesn't mean the work is done, there’s lots of work to do. But one of the things I want to be today is positive about what we can do. And, you know, I think it's important to recognise that because actually, if you just looked at the 2026 International AI Safety Report, you might go “Oh, they've kind of narrowed their focus, right?” The yellow ones are the ones that are in both reports, but the white ones are the ones that kind of dropped out of the 2026 report.

You might go, oh, those scientists are going into themselves. But of course, that's not true. That's what we have here actually is more of a division of labor. These conversations are expanding out and different groups are engaging in them. And so if you look at the international, the Independent International Scientific Panel on AI preliminary report, you'll see, the discussion of the science, as Liming has already pointed out. But there's also some really important work adding additional elements to this conversation about the AI, the challenges we have in AI governance.

So yes, we've got all the acknowledgment of accelerating capabilities, good and bad, but we also have recognition of one of the big challenges we're facing now, which is inequalities and asymmetries, right? Asymmetries in the opportunities, asymmetries in where the harms are falling asymmetries of information and not just between countries, but also between governments and firms. And that specifically called out in that report in really important ways. And it's a governance challenge that we need to face. If you dig further into the report, it highlights some of the key challenges that face us, when we do think about AI governance and policy, and as it intersects with safety.

You can see here they talk about something that I'm sure you've all heard, you know, technology races ahead and law kind of limps behind, eventually catching up. So that development is outpacing regulation. The challenge that Frontier Labs are withholding models or defining their own safety standards. I think we've heard some significant qualifications of that today. The governance approaches are fragmented. And then challenges around epistemic impacts and asymmetries of power, as I said. Just in the small amount of time I have left. I just want to respond to some of these challenges.

First, the question of developer development is outpacing regulation. And first, it's obviously true that democracies, and countries indeed struggle to write new law quickly. Making law and making legal decisions through courts and the like takes time. It does. But there is, in fact, a lot more regulatory action in this AI space than you perhaps think. We have, locally, Raymond Sun, who does this global AI Regulation Tracker. These aren't laws. Those numbers aren't laws, but they are regulatory type actions, bills, proposals, suggestions. And what we what you know, what he concludes from his really quite impressive review is that almost every jurisdiction now regulates AI in some way.

And I can recommend the website if you're interested. You can click on countries, look at what's been happening in each individual country. That's an impressive piece of work. And you can see he's actually giving a talk on Wednesday evening at UNSW, look him up on LinkedIn. So you just have to look beyond the big jurisdictions. We get very focused on the US, where not a lot happens at the federal level, but at the state level in the US and in countries all over the world, regulatory action is happening.

We have to find those, we have to work out what works. Second, I think it's important that you remember how law works, right? Law isn't just writing laws. Regulation and governance does not just come in the form of shiny new AI laws. Increasingly, courts all over the world are starting to look at aspects of AI practice and saying, “Yeah, you can be held responsible for that. That was you. That harm was you”. And we can hold you liable. So, we do have Meta and YouTube being found liable in the US for social media addiction by a jury, admittedly, but still.

We have the regional court, finding Google responsible for the AI overviews it produces. And these court cases create a kind of moment and an opportunity because it isn't just courts that enforce standards. Yes, it's taken a long time to get to those court decisions, but decisions like this feed into a broader network of enforcement. Insurance, professional standards, procurement standards, duties of company directors. As we start to see these cases, you are starting to create the conditions whereby those broader systems can start imposing incentives on AI firms and making it clear that there is, in fact, legal risk.

For a long time, we've operated as if there was no legal risk. I think that is changing, and that's an important point to recognise. Another important point to recognise is how that connects to your work on AI safety and standards and evaluation. Because as I say over there, the evidence base for liability is building. Every time we identify risks, develop methods to identify those risks, get evidence of harms, get evidence of what you could do to address those harms in advance, you are building not just technical standards, but in effect, standards of professional practice that can be the basis for legal standards and legal liability.

All of the work we are doing in AI safety connects to this. So there's that. There's another moment we are having, that is, of course, the Fable/Mythos moment, which is a different kind of moment. It's a moment that highlighted for us and for governments around the world, the risks of dependency and the reality of uneven access. And to me, that is a learning moment, both for us, but also for the frontier labs, that there could be active benefits in cooperation around safety, because the last thing you want is, you know, presidents

being arbitrary with the standards. So there's already been extensive work done that can be the basis for standards here. So we have these opportunities, opportunities, to take this moment and build governance at the same time as we are running to build safety standards and running to understand the technology. And there's a lot of people working on this, and there are lots of forums for that discussion. But there are a couple of core hard parts that I do want to highlight, and this is the stuff that's keeping me up at night.

I've tried to be upbeat, but let me be a little bit sad for a little bit for just a moment. Most of what I have been talking so far has been actually classic territory for law, and governance: identifiable direct risks, measurable harms law, and courts that can respond. Law and Policy find it a lot harder to address truly systemic challenges. Global technology divides are not something we have a great history of solving. Everyone in this room lived through COVID and the vaccine battles that came around that.

AI appears to be a technology that accelerates inequality and concentration of power. So one of the projects I'm working on, I'm thinking about a lot is around questions of interoperability, open innovation and how we keep things, not locked down, not concentrated, and make sure that there's still a role for smaller innovators here in Australia and elsewhere. The second thing that's keeping me up at night is what the international AI Safety Report calls autonomy risks, or what the UN report calls epistemic impacts and the challenge of the erosion of shared reality, what the Vatican calls preserving what about us is human.

The superpower at the heart of democracy lies in our ability to talk to each other, to hear each other, and to reach common ground. Platforms and social media have harmed that shared reality. AI threatens to harm it thoroughly and to call out the bad angels ourselves. So the big challenge we have is to address that. But I believe, fundamentally, I think we can address this because fundamentally, I believe in humans' abilities to connect, to take agency and to do right by others. It's not inevitable. That's where the work is.

That's where all of us in this room have to be working. So more power to us. Let's have a great couple of days and start doing that work.