Artificial Intelligence (AI) technologies are deeply embedded at the core of today’s online platforms, shaping how content is created, curated, and moderated. Large language and vision models now generate persuasive text and realistic images. Recommendation algorithms play a central role in deciding what people see and engage with online, while AI moderation tools filter vast streams of user-generated content. Despite this deep integration, our understanding of how these systems work and their broader societal impacts remains limited.

In this talk, Dr. Zannettou will present three data-driven studies that treat online platforms and AI models as auditable systems through the lens of robust and safe AI. First, he will talk about how prone modern text-to-image models are to producing harmful or policy-violating imagery across diverse prompt conditions. Second, he examines automated detection of hateful imagery with a focus on temporal robustness: particularly, how good are state-of-the-art AI models in measuring the evolution of harmful imagery? Third, Dr. Zannettou audits algorithmic transparency online using digital-twin-inspired sockpuppets (controlled proxies for user behavior), to test what platforms disclose versus what users experience in practice. He will conclude the talk with a discussion of my future research directions.

Dr. Savvas Zannettou is an Assistant Professor at the Delft University of Technology and an associated researcher at the Max Planck Institute for Informatics. His research applies AI and large-scale quantitative analysis to audit online platforms and understand how AI-driven systems (e.g., recommender algorithms) impact societal-level phenomena, including the spread of misinformation and hate speech.

His research has been published in top-tier venues such as WWW, ICWSM, ACM CCS, IEEE S&P, and Usenix Security, and has received multiple awards, including a Distinguished Paper Award at ACM IMC 2018 and Best Paper Honorable Mentions at ICWSM 2020 and 2024, CSCW 2021, and ACM CCS 2022.