During IJCAI-ECAI 2026 in Bremen, I had the opportunity to be interviewed by Elena Müller from Gesellschaft für Informatik (GI) about my research and some broader questions around reliable and trustworthy AI.

The conversation started from the work I was presenting at IJCAI on self-supervised learning and abstention for diabetic retinopathy screening. From there, we discussed a broader question that appears across many AI applications: how can we build models that not only make good predictions, but also recognize when they should not make a decision at all?

One of the main topics was the interaction between data-driven models and rule-based systems. While modern AI models are very powerful at learning patterns from data, explicit rules and constraints can provide an additional layer of control, especially in safety-critical applications. In healthcare, for example, a system could use such mechanisms to refer uncertain cases to a medical expert instead of forcing a prediction.

We also talked about AI reasoning, explainability, model efficiency, and the increasing use of AI in research workflows. An especially interesting part of the discussion was the role of human creativity as AI systems become more capable. While these systems can already support researchers in many ways, identifying meaningful new questions and deciding which directions are worth exploring still remains an important human contribution.

Another topic close to my research interests was the development of smaller and more efficient AI models. Increasing model size alone is unlikely to be a sustainable solution, particularly if we want capable AI systems to remain accessible without requiring increasingly large computational resources.

I really enjoyed having the opportunity to discuss these questions outside the usual format of a technical presentation and to reflect on how reliability, efficiency, and human oversight might shape the next generation of AI systems.

Many thanks to Elena Müller and Gesellschaft für Informatik for the conversation and for featuring the interview.

The full interview, “Rule-Based Models Should Be the Backbone”, is available on the Gesellschaft für Informatik website.