This course addresses the fundamental challenge of aligning AI with human values. We will explore classic and contemporary theories of value and determine why it is difficult to convey these values to a machine. The course covers the three main types of alignment failure, corresponding to the three main machine learning training methods: bias amplification (in supervised learning), hallucination (in unsupervised learning), and reward hacking (in reinforcement learning). Students will learn to identify, mitigate, and communicate about AI bias, examining real-world scenarios where the tolerance for bias differs, from medical diagnostics to image generation. The course will explore various fairness metrics, their limitations and trade-offs, and how to apply them effectively. By the end, students will be prepared to lead interdisciplinary teams responsible for assessing, explaining, and mitigating misalignment.
This course empowers future AI leaders to strategically identify and evaluate potential AI applications within their organizations. Students will learn to map business processes, contrast job descriptions with actual workflows, and pinpoint areas where AI can drive strategic advantage and enhance efficiency, productivity, or innovation. Additionally, the course will address strategies for integrating existing, potentially unsanctioned AI solutions into a centralized framework, ensuring compliance while harnessing the creativity and productivity gains of grassroots AI adoption. Throughout the course, we will emphasize aligning AI initiatives with overarching organizational goals and values, regardless of the sector (business, government, non-profit, etc.). As part of the coursework, students will conduct a strategic assessment of AI opportunities within a chosen organization or context, identifying potential use cases, evaluating their feasibility, and aligning them with the organization's strategic vision.