Social AI for Everyday Assistance

Program Description

AI systems are increasingly entering everyday life through systems situated in real-world physical and social contexts, such as autonomous vehicles that transport us or chatbots that provide emotional support. While these social AI systems often perform impressively, they also act inappropriately leading to breakdowns during interaction. Social AI presents serious design challenges, particularly as people increasingly rely on AI systems for everyday assistance. 

One illustrative example of social AI acting inappropriately can be related to sycophancy, a term that has recently regained attention as a consequence of AI systems over-prioritising agreement with user input, even when doing so may undermine safety, well-being, or task success. Such behaviour raises social questions about how to design socially appropriate AI and when systems should challenge or override user requests, as well as technical questions such as how appropriate sociality can be conceptualized ‘under the hood’. 

Addressing these challenges requires multidisciplinary collaboration between researchers focused on the cooperative work that engaging with social AI entails, and researchers with expertise and experience of technical challenges. At present, however, perspectives on these challenges largely develop in parallel with limited overlap.

The P1 program, Social AI for Everyday Assistance, aims to bridge these gaps by bringing together researchers with diverse knowledge and expertise, including but not limited to human–computer interaction, machine learning, linguistics, and natural language processing. 

The objective is to establish a platform for discussion that surfaces diverse perspectives and synthesizes them into a shared vision by (1) collecting and understanding community perspectives on Social AI and sycophancy; (2) reflecting on critical edge cases, such as blind individuals interacting with AI to cross the road; and (3) outlining concrete design guidelines to support future research on meaningful interactions with social AI systems.

With the ambition to engage researchers across disciplines to push our understanding of and envision future Social AI systems for everyday assistance, four events will be held, focusing on:

1. Interdisciplinary discussions exploring social and technical perspectives, approaches, and ways forward, while building and forming interdisciplinary relationships between program members
2. Social AI Edge Cases including hands-on practical work and discussions to unfold perspectives and push unification of approaches
3. Disseminating insights by organising a Social AI Symposium to revise and validate understanding of Sycophantic AI, and extend collaboration with international researchers
4. Formalizing a vision for Social AI