Reflections on XAI

About

Explainable AI (XAI) is now deployed widely, including in high-stakes domains such as medicine, and explanations are even required by law. Explanation methods must therefore reliably serve their purpose rather than offer convincing but incorrect explanations.

Over the past decade, XAI has grown into a substantial field, yet much recent work focuses on incremental technical questions rather than its original ambition: making AI systems meaningfully understandable, trustworthy, and useful. Meanwhile, foundation models and LLMs have changed the landscape, and it is almost ten years since Finale Doshi-Velez and Been Kim's call for a rigorous science of interpretability.

This session is an intermezzo for the XAI community: leading researchers critically assess what XAI has achieved in practice, where it has fallen short, and which fundamental challenges should define its next phase. We cover different paradigms, from classic explainability approaches to chain-of-thought and mechanistic interpretability, and perspectives ranging from machine learning theory and practice to philosophy of science.

The session is guided by the following overarching questions:

1. Has explainability research drifted towards self-imposed, incremental questions instead of its original ambitious goals?
2. Is XAI keeping up with the rapid development of ML, and specifically of foundation models?
3. Are there clear success stories, and what is preventing more of them?
4. After 5–10 years of work, where should we redirect our efforts?

Invited Speakers

Timo Freiesleben LMU Munich

Hadas Orgad Kempner Institute, Harvard University

Tim van Erven University of Amsterdam

Lesia Semenova Rutgers University

Schedule

Each talk is 20 minutes, followed by 10 minutes of discussion and Q&A.

Talk titles and the room will be announced later.

Call for Posters

We invite posters offering critical perspectives on explainability and interpretability: its foundations, practical impact, and future directions. All paradigms are welcome, from feature attributions to mechanistic interpretability, and from theory to philosophy.

What you can submit

• A paper already accepted elsewhere (a conference, journal, or workshop) that fits the theme, or
• An extended abstract of up to 4 pages (excluding references). Work in progress is welcome.

Selection

Submissions are evaluated for fit with the theme; if spots run short, we will select among them. Contributions are non-archival. Posters must be A1 portrait.

How to submit

If you want to attend, please register for the ELLIS UnConference through the official registration form.

Organizers

Maximilian Thiessen University of Tübingen

Emmanuel Esposito CREST, ENSAE, IP Paris

Gunnar König University of Tübingen

Ulrike von Luxburg University of Tübingen

Nicolò Cesa-Bianchi University of Milan

This session is a follow-up to the workshop Theory of Explainable Machine Learning at the ELLIS UnConference / EurIPS 2025 in Copenhagen.

Contact

For questions about the session or the call for posters, please contact Maximilian Thiessen at maximilian.thiessen@uni-tuebingen.de.