Designing AI systems

Designing AI systems

Stoimenova, N. (TU Delft Methodologie en Organisatie van Design)

Kleinsmann, M.S. (promotor)
Gonçalves, M. (copromotor)

Delft University of Technology

Delft University of Technology

2023-10-09

Traditionally, the AI field has focused on providing answers in domains like chess, translation, protein folding, and item recommendations. These AI systems have a clear objectives and outcomes that can be easily classified as correct or incorrect. However, once such systems face the complexity of social contexts, they start to produce unintended and sometimes harmful outcomes. If we are to design AI systems that avoid far-reaching harmful consequences, we cannot decouple/detach them from the complex systems in which they operate.

A key challenge in doing so is AI systems' behaviour-use interdependence (i.e., the behaviour of a system is related to the manner, in which it is used) – a topic underrepresented in extant literature. We set out to explore this issue, guided by the central research question: "How can we design a theoretical model that facilitates early simulation of AI systems' behaviour-use interdependence using Design theories?”.

This dissertation presents both theoretical and empirical investigations into the development of such a model. These lay the foundation for what we term the Theoretical Model for Prototyping AI or PAI model. The PAI model is defined by relationships among abduction, induction, and deduction, which offer a means to support the early simulation of the behaviour-use interdependence. Finally, the devising of the PAI model allows us to shed light into how Design theories could contribute to the design of better AI systems. It also allows us to extend these theories and identify potential future directions for the field of Design.

Abduction
AI systems
unintended consequences
Design Theory
Design Cognition
behaviour-use interdependence
Reasoning
Design Methodology

https://doi.org/10.4233/uuid:c53b6902-f347-4694-9702-6bc167fb8c04

978-94-6366-742-5

Institutional Repository

doctoral thesis

© 2023 N. Stoimenova