Axel Flinth: "AI models are rational – humans are not. How severe is this problem?"
Speaker: Axel Flinth, Dept Mathematics and Mathematical Statistics, Umeå University.
Title: AI models are rational – humans are not. How severe is this problem?
Abstract: AI-agents are capable of learning to copy many behaviors. An interesting approach to this is so-called inverse reinforcement learning, where the approach is to find a utility function that makes the behaviour ‘rational’ – that is, maximizing the utility. However, humans do not always take ‘rational’ decision when choosing between insecure outcomes. This insight has spurred the development of an entire subfield of economics, so-called *behavioral economics*. This begs the question: What are the potential consequences of training a ‘rational’ AI-model on ‘irrational’ behavioral data?
Axel Flinth is together with Jonas Westin starting up a project about mathematically treating this question. What behaviors can and cannot AI-models based on certain *decision theory frameworks* behave, and when can they be learned from data? His ambition is to give a light-hearted intro to the problems, some decision theory frameworks, and, if time allows, some initial results.
Researchers affiliated to WASP-HS are especially welcome! We believe that there are clear opportunities for collaboration here, in particular with the upcoming WASP/WASP-HS joint call in mind.
