| Hello, and welcome to this week’s installment of the Future in Five Questions. This week we interviewed Helen King, Google DeepMind’s senior director of responsibility. As one of Google’s top decisionmakers on artificial intelligence risk and governance, she discussed how far the AI community has to go in learning how to evaluate its own tools, how AI might influence human decision-making and why everyone needs to get on the same page about what, exactly, “red-teaming” is. An edited and condensed version of the conversation follows: What's one underrated big idea? To build AI systems that are safe, ethical and beneficial for everyone, we need to bring together points of view from fields that may not typically intersect. The first step should happen internally, by assembling multi-disciplinary teams. The next step is to engage with third parties who can add specialized knowledge to decision making, to introduce a position we might not have thought about. This was a core part of how we determined the release of our AI system AlphaFold. Through consultation with experts across biology, pharma, biosecurity and human rights we built an understanding of how our release strategy could balance benefits and risks. That wouldn’t have been possible if we didn’t think beyond the walls of Google DeepMind. What’s a technology that you think is overhyped? Right now, I’d say AI evaluations tools are getting attention, but it’s somewhat premature. Tools that help researchers identify capabilities and unwanted behaviors in AI systems will be crucial for managing risks and ensuring AI benefits everyone. But it's important to acknowledge how nascent the field still is, despite the enormous progress made. There are gaps in our collective understanding of how evaluations should work and what defines a “good” evaluation. One effort I’m particularly proud of is the launch of our Frontier Safety Framework, a set of protocols for proactively identifying future AI capabilities that could cause severe harm and putting in place mechanisms to detect and mitigate them. It’s also positive to see efforts such as the ML Commons, Google’s SAIF framework, and the Frontier Model Forum’s AI Safety Fund helping fill the funding gap by providing grants to independent researchers exploring some of the most critical safety risks associated with frontier AI systems. What book most shaped your conception of the future? A book I read many years ago and often come back to is a book called “Yes!” by Noah J. Goldstein, Steve J. Martin, and Robert B. Cialdini. It’s about humans, psychology, and persuasion and contains food for thought on how we are influenced and persuaded. I often think about how that will translate to the future as generative AI systems become increasingly advanced and could influence decision-making, and it’s something my colleagues are thinking about too. What could the government be doing regarding technology that it isn’t? Consistency in the way that different companies and countries are evaluating the capabilities of AI systems is essential. To achieve that, we all need to be on the same page when we are talking about specific approaches to evaluations. “Red-teaming” is a very valuable practice, but different parties have different understandings of what that means, so it’s important for us to have a rigorous discussion about that before we jump to over specifying requirements in legislation. Governments can play an important role here by working closely with industry, civil society, and academia to agree on what specific terms mean, and help drive consistency in the way we use them. This in turn helps us set the right safety norms. What has surprised you the most this year? I’ve been working at the forefront of AI development and research for over a decade, and have been pleased to see so much collaboration across AI safety around the world — especially during the past year. It’s taken a lot of different forms, from the Frontier AI Safety Commitments bringing together 16 leading AI companies to agree on specific steps to take to develop AI safely, to the launch of dedicated AI Safety Institutes across 10 countries and the European Union. The Institute for Advanced Study also brought together scientific experts across safety and ethics to align on guidance for evaluation. This type of collaboration is essential to help us develop a shared understanding of the opportunities and risks presented by frontier AI models, and ensure billions of people around the world benefit.
| 
