Foundations of Responsible AI: Principles and Tradeoffs

Course Info

This course develops the mathematical and algorithmic foundations of responsible AI. We will study how to formalize concerns about privacy, fairness, transparency, and automated decision-making, and examine when these goals can be achieved together and when they conflict. Students will learn to design and audit AI systems and connect technical guarantees and limitations to ethical and policy questions.

Prerequisites

Probability, linear algebra, Python programming, and mathematical proofs. Familiarity with statistics and machine learning is helpful but optional. This course is an elective.

Course Topics

• Algorithmic fairness and decision-making. Sources of bias, group and individual fairness, calibration, multigroup fairness, auditing, causal models of discrimination, counterfactual fairness, and incompatibility results.
• Privacy and memorization. Reconstruction and membership inference, differential privacy, private learning, and tradeoffs among privacy, utility, and fairness.
• Strategic behavior and dynamic effects. Strategic classification, gaming, selective labels, feedback loops, performative prediction, and how deployment changes future data.
• Transparency, explanation, and accountability. Interpretable models, post-hoc and counterfactual explanations, algorithmic recourse, limits and stability of explanations, and model auditing.
• Selected topics in modern AI. Possible topics include LLM memorization, mechanistic interpretability, evaluation gaming, reward misspecification, and alignment.

Learning Outcomes

• Formalize concerns about AI systems by identifying stakeholders, harms, assumptions, objectives, and constraints.
• Apply and compare formal definitions of responsible AI goals, including in strategic and dynamic settings.
• Analyze canonical algorithms and interpret their guarantees.
• Explain tradeoffs and impossibility results.
• Design and evaluate a simple algorithm or audit.
• Communicate technical results and their practical, ethical, and policy implications through analysis, discussion, and writing.

Grading

• Assignments (50%). Approximately four assignments combining mathematical analysis with short written analyses of formal models. One includes a programming or auditing component.
• Preparation and discussion (10%). Brief responses to selected readings, case studies, or policy texts, followed by class discussion and activities.
• Tests (40%). Two tests worth 20% each, each covering approximately half of the course.

Course Materials

See Canvas for course materials and key dates.

Recommended Books

There is no required textbook.

• Solon Barocas, Moritz Hardt, and Arvind Narayanan. Fairness and Machine Learning: Limitations and Opportunities.
• Cynthia Dwork and Aaron Roth. The Algorithmic Foundations of Differential Privacy.
• Ferdinando Fioretto and Pascal Van Hentenryck, editors. Differential Privacy in Artificial Intelligence: From Theory to Practice.
• Christoph Molnar. Interpretable Machine Learning: A Guide to Making Black Box Models Explainable.
• Mehryar Mohri, Afshin Rostamizadeh, and Ameet Talwalkar. Foundations of Machine Learning.
• Shai Shalev-Shwartz and Shai Ben-David. Understanding Machine Learning: From Theory to Algorithms.

Academic Integrity and AI Use

Complete graded work independently except where the instructor specifies group work. You may discuss assignment specifications with classmates, but you may not discuss solutions. Cite any outside code or help that the instructor permits and list all participants on group projects.

You may use generative AI to clarify concepts or explore high-level ideas. You may not use it to generate assignment solutions, code, or written analyses, or ask it to answer or outline responses to assignment questions. Complete assignments independently after using a tool. Disclose which tool you used and briefly describe how you used it. You are responsible for your submissions.

University academic integrity policies are available for undergraduate students and graduate students.

Schedule

Schedule: TBD.