Despite the growing adoption of machine intelligence, one significant limitation remains a major risk factor. More specifically, machines are limited in their ability to understand and serve humans. This shortcoming looms behind current concerns about AI misalignment, AI misuse, and generally missed opportunities in developing AI that better collaborates with human partners. Current approaches to machine intelligence lack the necessary machinery to achieve this.

This Dagstuhl Seminar on Computational Theories of Interactive Behavior will bring together experts from cognitive science, AI, and human factors to advance a unified computational understanding of human thinking and behavior for interactive systems. Instead of asking how to build guardrails for AI post hoc, we ask how to build models that reliably represent and learn from human behavior. Novel machine learning methods allow the expression of theory-informed priors and structures. Such models are essential for reducing the risk of misalignment between AI-driven systems and human goals.

The main objective of this seminar is to form a new foundation for computational theories of interactive behavior for machine intelligence. To develop computational theories of interactive behavior that predict and explain how users interact with technology in everyday environments, we have formulated four primary goals to be achieved during the seminar, each deeply interrelated and pursued in an integrated manner. First, we aim to identify and reach a consensus on the key modeling challenges across individual disciplines (Goal 1). Next, we will discuss how a unified architecture of computational cognitive models can help to address these challenges and why it is necessary to agree on it (Goal 2). Building on this foundation, we will define a research roadmap to guide the development toward such a unified architecture (Goal 3). Finally, we will establish a joint curriculum to train the next generation of researchers in computational cognitive modeling of interactive behavior (Goal 4).

Creative Commons BY 4.0

Birsen Donmez, Patrick Ebel, Antti Oulasvirta, Philipp Wintersberger, and Jussi Jokinen