ELIAS & ELLIOT Theme Development Workshop on AI for a Sustainable Society

Communities should be central actors to imagine, design, and use AI. The talk will present recent experiences with participatory AI approaches — in which stakeholders share their knowledge and experience to work on problems that matter in the context of cities and education — and argue that democratic engagement of the public is fundamental for a sustainable future.

Daniel Gatica-Perez leads the Social Computing Group at Idiap Research Institute and is a professor at EPFL in Switzerland. His work integrates human-centered and participatory methods with mobile, social, and AI technologies to study how people and technology intersect in everyday life, and to support individuals and communities using these technologies. He also works with organizations on social innovation projects.

Foundation models are often presented as general-purpose systems that can transfer across languages, images, and contexts. Agentic systems may extend these capacities through increased inference time, retrieval, web access, and tool use, but these are additional components and not inherent properties of the model. This talk examines where claims of generalisability reach their limits.

Examples from language, vision, and spatiotemporal reasoning show how models reproduce Anglocentric and culturally specific assumptions. The problem is not only missing data, but also the categories and evaluation standards built into these systems. The talk therefore distinguishes scalable diversity, adding more languages, modalities, or domains to pre-defined structures, from meaningful diversity, which by definition changes things and requires epistemic nuance, plural standards of competence, and accountability to relevant communities.

Paula Helm is Professor for Empirical Computational Ethics at the Center for Critical Computational Studies, Goethe University Frankfurt. Her work brings together Critical AI Research, Empirical Ethics, and Science and Technology Studies. She studies computational systems not only as technical artefacts, but as infrastructures through which values, power relations, and forms of knowledge are organised.

Before joining C³S, Helm was Assistant Professor of Ethics and Data Science at the University of Amsterdam, where she served as Scientific Lead of the Empirical Ethics Research Group at the Amsterdam Institute for Advanced Studies and co-coordinated the MA track Cultural Data & AI, which combines critical inquiry with data science.

A central concern in her work is how AI ethics can move beyond abstract principles and public-facing commitments toward reflexive empirical inquiry, infrastructural accountability, and more just forms of technological design.

Paula is an ELLIS scholar and regularly serves for and publishes in interdisciplinary venues like FACCT, Big Data & Society or the Cambridge Forum for AI.

Models internalize more about their training data than their predictions reveal. We present BEE, which exposes spurious correlations through the weight space and embedding geometry underlying model decisions, and Semantic Correlation Descriptors, which turn such learned correlations into dataset specific fingerprints for dataset level membership inference. Together, these works connect model explainability with training data provenance, showing how internal semantic structure can reveal both harmful shortcuts and traces of the data that induced them.

Elena Burceanu is a Principal Research Scientist at Bitdefender, where she works on robust and reliable machine learning, with a focus on OOD generalization and spurious correlations. She holds a PhD from the University of Bucharest, Romania and is actively involved in the ML research community, including serving as an Associate Program Chair at NeurIPS 2025, Local Chair for CoLLAs 2026, and co-organizing EEML summer school and RoAIDays in Bucharest. Her recent work focuses on understanding what models learn from their data, why unintended correlations emerge, and what these learned patterns can reveal about the training data itself.

Frontier AI models are rapidly transforming cyber offense and defense by discovering vulnerabilities, drafting exploits, and generating patches. These dual-use capabilities undermine classic models of vulnerability disclosure, which assume discovery is scarce and costly. So when a lab releases a frontier model, who gains the upper hand: defenders or attackers?

We present The Oracle’s Gambit, a three-player leader–follower (Stackelberg) game between frontier AI labs, software vendors, and adversaries. We study how release policies, from full open release to defender pre-release windows and embargoes, shape the downstream race between patching and exploitation.

Security depends on the capability gap a pre-release opens for defenders, not on capability level alone. Symmetric access can intensify attacks, while a timed pre-release gives defenders a temporary edge, and it is the optimal lab policy in 80–95% of simulations under parameter uncertainty. Release timing is therefore a concrete governance lever for protecting the digital infrastructure society depends on.

Christoph Landolt is an ELLIS PhD student at CISPA Helmholtz Center for Information Security, supervised by Mario Fritz. He works on agentic AI for cybersecurity, focusing on multi-agent learning and game theory. Previously, he was a research fellow at the Swiss Cyber-Defence Campus, training autonomous penetration-testing agents, and a machine learning research assistant at the Eastern Switzerland University of Applied Sciences (OST).

Market mechanisms shape how goods and resources are exchanged. AI systems supporting these decisions must often learn from limited observations, without access to participants’ private valuations. In this talk, I will discuss the sample complexity and regret of learning fixed-price mechanisms, highlighting connections to learning multidimensional cumulative distribution functions.

Matteo Castiglioni is an assistant professor in the Department of Electronics, Information and Bioengineering at Politecnico di Milano, where he also received his PhD in Computer Science. His research focuses on learning and decision-making in complex multi-agent environments, combining approaches from machine learning, algorithmic game theory, and theoretical computer science.