In an era marked by the rapid advancement of artificial intelligence (AI), the intersection of philosophy and AI research has become increasingly vital. In this Dagstuhl Seminar, we single out three clusters of pressing issues related to AI research—foundation models, causal inference, and Bayesian learning—that could benefit from a philosophical inquiry across its subfields—philosophy of mind, philosophy of science, and epistemology.

Foundation models have been central to recent advancements in AI. Even though at their core these models just predict the next token in a sequence, beyond certain thresholds of data and computational power, they unlock new capabilities. Are these capabilities indicative of understanding in virtue of, for example, possessing a world model, or do models remain mere “stochastic parrots”? Additionally, can foundation models be conscious, and how could we measure it from the outside? The philosophy of mind has long dealt with the elusive nature of understanding and consciousness. Additionally, beside long-standing philosophical positions on the nature of consciousness, recent work looks at neuroscience as well, drawing lessons for the possibility of artificial consciousness.

While machine learning excels at identifying correlations within vast datasets, it falls short in terms of identifying true causal relationships, which are crucial for many applications, particularly in fields where the stakes of intervention are high. Despite its appeal, the challenge of causal discovery from merely observational data remains daunting, raising fundamental questions about its feasibility. What kind of causality is needed for practical applications of AI, and which conditions must a dataset meet for causal learning to be meaningful? The philosophy of science has recognized the multifaceted nature of causality related to inference, explanation, prediction, control, and reasoning. The range of causal concepts it showcases and analyses of when non-experimental methods are sufficient for knowledge generation can be a useful resource in these matters.

Finally, Bayesian inference, with its core principle of marginalizing over all possible explanations for the observed data, offers highly flexible probabilistic models with many desirable properties. Many mathematical arguments for the optimality of Bayesian learning often assume exact inference in well-specified models, while practical applications of machine learning are confronted with the challenge of model misspecification and the intractability of exact inference. This necessitates the use of approximate methods, raising the question of whether approximate Bayesian inference retains the theoretical benefits of its exact counterpart. More fundamentally, are probabilistic uncertainties modeled by Bayesian methods the right kind of uncertainties for real-world machine learning problems? Bayesian epistemology has long grappled with the normative requirements of the framework and the ways to make it less idealized and more realistic, inquiring when it is better than its non-Bayesian rivals.

To address these intellectual challenges, the seminar wants to bring together a unique blend of leading experts in AI and philosophy, with a focus on the areas discussed above. Through a series of activities, including short introductory lectures, speed networking sessions, interactive discussion sessions, and working groups, we aim to foster a practice-oriented interdisciplinary collaboration. By connecting these two formerly disjoint communities, the progress made over the course of this seminar will lay the groundwork for a more prudent approach to scientific inquiry in the rapidly advancing field of AI, anchoring the research to a solid foundation of careful thought for the years to come.

Creative Commons BY 4.0

Alexander Bird, Vincent Fortuin, Bojana Grujicic, and Mohammad Emtiyaz Khan