Why do you care about AI Existential Safety?

I care about preserving human agency and our ability to shape the future. AI could help us make enormous progress, but we are building increasingly capable systems without a comparable understanding of how they work or when they might fail. As these systems take on more consequential responsibilities, that gap becomes harder to accept.

My concern is that strong performance on a set of tests does not necessarily mean a system will behave safely in a new situation. We need to understand failures that could lead to a loss of control: systems pursuing unintended goals, concealing their actions or acting beyond their authorisation in ways that undermine our ability to monitor and intervene.

I also care about who has the ability to scrutinise and control these systems. Technical safety needs to support meaningful oversight, including independent evaluation and evidence that policymakers and the public can assess. I want my research to help us retain control of advanced AI and ensure that its development supports human freedom and flourishing.

Please give at least one example of your research interests related to AI existential safety:

A central question in my research is whether understanding a model’s internal mechanisms can help us evaluate and control behaviour that is difficult to detect through testing alone.

For example, my group studies mechanistic interpretability: identifying the representations and circuits that contribute to a model’s behaviour, then testing those explanations through interventions. I am particularly interested in whether these methods can help us detect deception and understand how situational awareness affects a model’s behaviour, especially when failures could undermine oversight or lead to a loss of control. Finding a plausible explanation is only a first step; we also need evidence that it identifies what actually causes the behaviour and remains useful beyond the examples we initially studied.

A related interest is understanding what a model’s chain of thought can tell us about its reasoning. A fluent explanation is not necessarily a faithful account of how an answer was produced. My work asks when reasoning traces help an overseer detect failures, what they leave out, and how to evaluate their usefulness for a particular oversight task.

Across these projects, my aim is to produce evidence that helps us make better decisions about advanced AI: what a system can do, where our understanding is limited, and whether proposed safeguards actually work. This connects my technical research with my work on AI governance and independent evaluations.