Thinking Globally, Testing Locally: Lessons in AI Safety Evaluation from the Polish Perspective

The Data Science Discipline of the School of EECS is hosting the following guest seminar:

Thinking Globally, Testing Locally: Lessons in AI Safety Evaluation from the Polish Perspective

Speaker: A/Prof Szymon Lukasik, AGH University of Krakow, Director of AI Safety Research Center, NASK National Research Institute, Poland

Abstract: The rapid advancement of artificial intelligence, particularly large language models (LLMs), has opened new frontiers across scientific, social, and economic domains. As these systems become increasingly integrated into everyday life, the imperative to ensure their safety, robustness, and alignment with human values grows stronger. While global efforts in AI safety research have made significant strides—addressing model alignment, hallucinations, robustness to adversarial inputs, and more—there remains a critical gap: the incorporation of local linguistic, cultural, and societal contexts into evaluation and development practices.

This talk explores what it means to evaluate AI safety locally, focusing on the Polish context. We will examine how the assumptions and benchmarks of global AI safety research must be adapted to capture real-world risks in underrepresented languages. We will showcase recent efforts in assessing the safety of the Polish Large Language Universal Model (PLLuM), focusing on both linguistic and sociocultural evaluation strategies. Additionally, the talk will cover techniques for identifying synthetic content—an increasingly pressing concern in information ecosystems. These case studies serve as a foundation for advocating more inclusive and locally aware AI safety research.

Speaker's bio: Szymon Lukasik is an associate professor at the AGH University of Kraków, Poland, and Director of AI Safety Research Center at NASK Polish National Research Institute. He obtained his PhD (2012) and habilitation (2019) in the fields of data analysis and computational intelligence and gained professional experience as a visiting scientist at the University of California Berkeley, UNINOVA (Portugal), National Laboratory of Pattern Recognition (China), and the Australian National University (Australia). He has contributed to 80 peeer-reviewed papers. His research interests cover nature-inspired optimization and multimodal machine learning. A member of IEEE, and IEEE Computational Intelligence Society, expert for the Polish National Centre for Research and Development and the European Commission. He serves as a reviewer for several journals in the field of data science and artificial intelligence.