The AI and Society (AISOC) team is is headed by Prof. Bilal Zafar. We are part of the Faculty of Computer Science at Ruhr University Bochum and affiliated with the Research Center for Trustworthy Data Science and Security (RC Trust), and the Cluster of Excellence CASA.

Our work aims to advance the development of trustworthy Artificial Intelligence (AI) and Machine Learning (ML) systems. By integrating multi-disciplinary perspectives, we aim to establish a comprehensive and practical framework for AI/ML trustworthiness. The key questions guiding our current research are:

• Quantifying LLM’s ability to understand the world: The performance of LLMs regularly matches or even exceeds humans on popular benchmarks. However, proficiency in solving benchmarks does not equate human-level reasoning. For instance, a high score on a graduate-level math benchmark rarely implies the same level of conceptual understanding as a graduate-level math student. Since LLMs understand the world very differently than humans, our work aims to quantify these discrepancies.
• Measuring and calibrating trust in AI: Despite their impressive capabilities, AI models continue to exhibit vulnerabilities like bias, lack of faithful explanations and hallucinations. At the same time, it is not always clear what it means for a LLM to be biased and for an explanation to be satisfactory. Our work draws insights from social sciences to operationalize these concepts and mitigate problematic behaviors.
• Deployment challenges: Current day AI models are so compute-intensive that efficient deployment requires specially designed acceleration procedures like weight quantization and token pruning. These strategies, while enhancing efficiency, can have an adverse impact on the predictive performance. Our work aims to understand the impact of acceleration procedures and balance efficiency with performance.

For more details, check out the full list of our publications.