Artificially intelligent (AI) systems are playing an ever increasing role in society. To ensure that they are beneficial, we must guarantee that they act not only according to the designer's intentions, but also in a way that is fair to all individuals and groups affected by their decisions. Many applications integrate human and AI decisions, such as crowdsourcing, recommendation systems, navigation and autonomous vehicles, as well as decision support systems for credit risk and criminal recidivism. AI systems are also deployed in semi-automated design tools, and diagnostics in medicine. Humans interact with a pre-designed AI, which may by unaware of the motivations, behaviour or knowledge of people. When such an AI is used at scale, it must take into account broader, societal considerations, such as fairness and privacy. Our project will investigate the relationship between humans and AI in society. We will create algorithms for automatically aligning AI behaviour with societal values. This involves learning about human preferences and behaviour, and using this information so as to take both into account. Our work will be grounded on technical advances in reinforcement learning, mechanism design, differential privacy, and the emerging field of algorithmic fairness. A novel aspect of our proposal are informational notions of fairness, where the decision maker must consider feedback effects, and our explicit handling of human biases. The project will be implemented by the PI, one postdocotral researcher and two PhD students, in collaboration with researchers abroad. The latter will be enhanced through visiting researcher grants of 1-2 months per year. We expect that these will be mainly taken advantage of by the students hired by the project to travel to Harvard (USA), MPI (Germany) and EPFL and St. Gallen (Switzerland) where the PI has collaobrations in the area.
Project leader: Christos Dimitrakakis
Institution: Institutt for informatikk