Algorithms for Social Systems: Beyond i.i.d. Environments

Algorithms and AI systems are increasingly deployed at scale in consequential social settings, e.g., online platforms, labor markets, healthcare systems, and educational institutions. The standard methods used to design and evaluate these algorithms were developed under assumptions of independent and identically distributed (i.i.d.) data, assumptions that are systematically violated when humans interact with each other and respond to the systems around them. Researchers working in these settings need to understand when these assumptions break down, what the resulting effects on system performance are, and how to design better systems. This course examines the evaluation, deployment, and human response to algorithms in social systems where these i.i.d. assumptions fail. We study how interference between users, capacity constraints, and behavioral responses to algorithms undermine both classical machine learning methods and causal inference tools like A/B testing and RCTs — and for each failure mode, we examine recent methodological advances designed for these settings. Topics include experiment design under interference, recommendation systems and their downstream effects, human compliance and discretion in algorithmic decision-making, and the operational realities of deploying algorithms at scale. The course closes by applying this analytical lens to large language models, examining whether the same failure modes arise, and whether the same advances apply, as AI interventions scale across social systems. The class will draw on recent theory and methodological work in causal inference, operations research, and human-algorithm interaction to prepare students for research at this frontier. This course is targeted at mathematically mature PhD students in operations research, statistics, computer science, economics, business, or quantitative fields interested in the intersection of algorithms and human behavior.