Seminars and Events
Algorithmic Integrity and the Predictability Limits of Human-Centered AI Systems
Event Details
The explosive growth of data and computing power, coupled with advances in machine learning models, has created new opportunities for AI algorithms to improve the human condition. However, unthinking reliance on algorithms can mislead even the best-intentioned applications of AI in socio-technical systems. In this talk, I highlight some of the challenges to the threats of the integrity of AI. First, I show how heterogeneities endemic in data can lead traditional statistical methods astray, illustrating with case study of Covid-19, where aggregating statistics from heterogeneous communities can exaggerate estimated disease growth rates. Second, I discuss how partial observability of a real-world system limits the accuracy of forecasts of its behavior. I quantify the loss of predictability and show that it cannot be recovered with external data regardless of the forecasting model adopted. Finally, I discuss how we can understand the feedback loop between people and algorithms within socio-technical systems to understand the stability of algorithmic decisions and prevent unexpected outcomes. My talk highlights algorithmic integrity as a fundamental component of trustworthy AI systems.