Do AI systems inherit the same appearance-based biases as humans? A large-scale study with 2,700+ participants showing that the well-known attractiveness halo effect — the tendency to attribute positive qualities to attractive faces — persists and amplifies in AI-mediated judgment contexts, including beauty filters and automated scoring systems.

A cross-disciplinary framework that systematically maps cognitive biases from psychology and behavioural economics onto AI system design decisions. BIASeD provides AI designers with principled guidance for incorporating human irrationality into automated systems — turning biases from liabilities into design inputs.

Appearance-based discrimination (lookism) is pervasive in human society — but how deeply is it encoded in AI? This project evaluates lookism across computer vision models and text-to-image generative systems, developing evaluation benchmarks and surfacing structural biases in how AI systems represent and judge human appearance.

Trust in AI assistants is often studied normatively — what should drive trust? This project contrasts normative principles with behavioural drivers: the psychological mechanisms that actually govern how users form, maintain, and break trust with conversational AI systems, with implications for design and deployment of LLM-based assistants.

How does task complexity shape the performance of individuals and groups working within human-machine team structures, even without direct communication? Computational models of group behaviour were developed and validated at the CMU Dynamic Decision Making Lab to characterise performance bottlenecks in collaborative human-machine systems.