Title: When AI Sees Too Much: Responsive and Responsible Multimodal AI for Social Good

Abstract: Multimodal AI is increasingly able to infer subtle human states from text, video, behaviour and physiological signals. These capabilities create powerful opportunities for social good: safer transport, better healthcare communication, improved training, more accessible public services, and systems that can respond to stress, doubt, engagement, trust or misunderstanding. However, the same signals also expose sensitive information that people did not explicitly choose to disclose. This talk argues that multimodal content analysis for social good must be both responsive and responsible: responsive in using human signals to support people, and responsible in limiting what is collected, retained, shared and inferred. I will discuss examples from affective computing, physiological sensing, empathy detection, avatar trust, driver state monitoring, deception/doubt detection and privacy-preserving video and biosignal analysis. The central claim is that the next phase of socially useful AI is not just better models, but better judgement about which human-state inferences are valid, useful, permissible and accountable.

Title: Machine Learning for Social Good: Challenges with Vulnerable Populations

Abstract: The scale, reach, and real-time nature of the Internet is opening new frontiers for better serving humanity, and particularly for a deeper understanding of the vulnerabilities in our societies, including inequalities and fragility in the face of a changing world. Vulnerable populations including children, elderly, racial or ethnic minorities, socioeconomically disadvantaged, underinsured, or those with certain medical conditions, are often absent in commonly used data sources. The aim of this talk will be to shed light on data and algorithmic biases, as well as methodological particulars when addressing vulnerable populations, drawing insights from case studies in response to real humanitarian crises.