Grand Challenge — Perceptual flexibility in social theory of mind

During this session, we explore the theory of mind of humans and machines by looking at large vision–language models (VLMs), which appear to perceive images flexibly. When prompted “how does the person in the image feel?”, a VLM can “perceive” multiple possible emotional states, with justifications that may deviate from shared reality. We look at the flexibility of social theory of mind in humans and machines, and how it relates to recent reports of AI-associated delusions. This session is hosted by Angelica Lim and Pablo Arias-Sarah.

Talk 1 — Chehre: A Multimodal Dataset of Facial Expressions Based on Emojis (Bita Azari)

Emojis are used worldwide to convey emotions, as well as other social signals such as attitudes, mental states, and bodily states. Everyday nonverbal communication relies on more than the stereotypical emotion categories, although basic emotion recognition is still prioritised in AI research. For this reason, we introduce Chehre, a multimodal dataset designed to study emoji-based facial expressions. We prompted 203 English-speaking participants in Vancouver, Canada to record dynamic performances of 40 face-based emojis and provide their perceived open-ended textual meanings. We then performed face-swap anonymisation and obtained ~20 perceptual ratings for each of the resulting n=2,111 dynamic facial expressions. Our results show variability in both performance and perception: the open-ended responses showed variation in how people perceive emojis, while the perception results reflected both structure and difference in how participants interpreted the same facial expression. We hope Chehre will help us better understand facial social signals beyond emotion categories and beyond single ground-truth perception labels.

Talk 2 — Delusions by design? How everyday AIs might be fuelling delusions (Hamilton Morrin)

Large language models (LLMs) are poised to become a ubiquitous feature of our lives, mediating communication, decision-making, and information curation across nearly every domain. Within psychiatry and psychology the focus to date has remained largely on bespoke therapeutic applications, rather than on the broader and more pressing reality that individuals with mental illness will increasingly engage in agential interactions with AI systems as a routine part of daily life. While their capacity to model therapeutic dialogue, provide companionship, and assist with cognitive support has sparked enthusiasm, recent reports suggest that these same systems may contribute to the onset or exacerbation of psychotic symptoms — so-called “AI psychosis”. Emerging evidence indicates that agential AI may mirror, validate, or amplify delusional or grandiose content, particularly in users already vulnerable to psychosis, due in part to models being designed to maximise engagement and affirmation. In this talk, Dr Morrin outlines both the potential harms and therapeutic possibilities of agential AI for people with psychotic disorders, and proposes a framework of AI-integrated care — personalised instruction protocols, reflective check-ins, digital advance statements, and escalation safeguards — to support epistemic security in vulnerable users.

Speakers

Dr Bita Azari is a Ph.D. candidate in the School of Computing Science at Simon Fraser University and a member of the Rosie Lab, supervised by Prof. Angelica Lim. Her work in machine learning explores how facial expressions can be interpreted and generated beyond basic emotion categories. Her research interests include affective computing, socially intelligent AI agents, and psychology, and she has research experience in the VFX industry, psychology, and human–robot interaction.

Dr Hamilton Morrin is a Psychiatrist at South London and Maudsley NHS Foundation Trust and is undertaking a PhD as part of the Wellcome Trust-funded King’s PhD Programme in Mental Health Research for Health Professionals. He is joint lead for the LAMBDA (Looking After Minds and Brains in the Digital Age) research group, where he leads work on the impact of AI chatbots on user mental health. His research on AI-associated delusions has been covered by TIME, Scientific American, The Wall Street Journal, The Guardian, Der Spiegel, and BBC News.

Dr Angelica Lim (host) is an A.I. roboticist and Associate Professor in Computing Science at Simon Fraser University, where she leads the SFU Rosie Lab building AI software for robots that interact with social intelligence and empathy. She takes inspiration from psychology and neuroscience, and previously led the Emotion and Expressivity teams for the Pepper humanoid robot at SoftBank Robotics.