Cognitive scientist studying reasoning and representation in human and machine minds.

My work adapts methods from cognitive neuroscience — controlled experiment, functional localization, and targeted ablation — to ask what AI systems actually represent, and why they fail when they do. Recent work found that frontier language models track the right variables in mechanical-reasoning problems but lean on a counting heuristic, and fail when success requires representing how a system is connected.

I am now extending this programme to multi-agent safety: detecting collective goal drift in populations of AI agents, and asking how training environments can select for prosocial behavior.

I trained in the humanities and evolutionary anthropology (MSc, Oxford), then in cognitive science (PhD, Radboud), and was a postdoctoral researcher at Emory University, where I remain a visiting researcher. Along the way I co-founded two applied-AI companies, Texture AI and Ogma Therapy; Ogma was acquired by Saltroad in 2026. I live in Oxford.