Dagstuhl-Seminar 28061
XAI Meets GenAI: From Explainable AI to Explaining AI
( 06. Feb – 11. Feb, 2028 )
Permalink
Organisatoren
- Elena Cabrio (Université Côte d’Azur - Sophia Antipolis, FR)
- Kary Främling (University of Umeå, SE)
- Barbara Hammer (Universität Bielefeld, DE)
- Henning Wachsmuth (Leibniz Universität Hannover, DE)
Kontakt
- Michael Gerke (für wissenschaftliche Fragen)
- Christina Schwarz (für administrative Fragen)
Explainable AI (XAI) studies how to make decisions and behavior of AI transparent, interpretable, and ultimately understandable. Classic XAI techniques aim to explicate how different aspects of the output of an AI depend on different aspects of the input of the AI. However, such XAI has been argued to be mainly helpful for the development and debugging of AI methods rather than for users interacting with AI in their everyday life and work environments. In addition, multiple research lines have pointed out that there is not the one explanation that fits all user needs.
Generative AI (genAI) methods create new content in terms of text, images, video, and more. To date, the most prominent genAI technique are large language models (LLMs) that primarily focus on text, though modalities are becoming increasingly intertwined. LLMs have revolutionized numerous private and professional activities – not only because they output reasonable and often impressive answers to nearly any prompt, but also because they provide, for the first time, an intuitive interface by which anyone can interact with AI in natural language. A pervasive use case of LLMs is to provide explanations to users’ requests across different social contexts. Thereby, LLMs open up room to move from explainable to explaining AI, that is, to give explanations that actually fit the background and needs of the user. Still, current LLM explanations tend to lack critical properties such as factuality, consistency, audience fit, and – in consequence – trustworthiness. We expect that genAI capabilities need to be integrated with established XAI techniques in order to achieve substantial progress on such issues and, thus, toward making AI truly explainable.
This Dagstuhl Seminar will focus primarily on the technical opportunities and challenges arising from genAI-based explanations, while also accounting for the human side, including linguistic, ethical, and social aspects as well as goals and needs in practice. Relevant questions include:
- What kinds of data to explain with genAI? For what use cases and applications?
- How should genAI explain? What context to model, modalities to use, and roles to take?
- How to enhance XAI with genAI? How to technically realize their integration?
- How to evaluate genAI-based explanations? How to establish meaningful benchmarks?
- Is GenAI the missing link to true XAI? In what regards, and what is missing?
To progress on these and related questions, the seminar aim to bring together leading and aspiring junior researchers from machine learning, natural language processing, and core AI across the globe who have academic, non-profit, or industry backgrounds, along with selected experts from non-technical disciplines, such as linguistics, philosophy, psychology, and the behavioral sciences.
Henning Wachsmuth, Elena Cabrio, Kary Främling, andBarbara Hammer
Klassifikation
- Computation and Language
- Computers and Society
- Machine Learning
Schlagworte
- Explainable AI
- Generative AI
- Large language models
- Explaining
- Human-AI interaction

Creative Commons BY 4.0
