Fall 2025 Course Topic: Explainable AI

This class will introduce attendees to some of the factors driving the renewed interest in the field of eXplainable AI (XAI). The class will also introduce a broad range of techniques used to render AI "explainable." As AI becomes ever more integrated into a broad range of human activity, policymakers have caught notice. For example, GDPR regulations grant data subjects the right to know about, "...the existence of automated decision-making... [and] meaningful information about the logic involved, as well as the significance and the envisaged consequences of such processing." But what is "meaningful information"? Once determining that answer, how can one present that meaningful information so that the intended humans can understand it?

This course will begin with paper reading from seminal works in the XAI field to drive discussions on topics such as:

• What factors impact humans' ability to understand explanation?
• How should we evaluate if an explanation is good?
• Should the system be fully automated OR rely on a human-in-the-loop?
• If there is a human in-the-loop, what responsibilities should they have and how should they accomplish such tasks?
• Is it sufficient to treat the model as an opaque box OR should we make a new model that is more inherently explainable?
• What are some techniques to provide opaque box explanation?
• What are some strategies to make a model more inherently explainable?
• What is the relationship between software engineering testing methodologies and XAI?

Once the class has attained some baseline knowledge through reading and discussion, we will move into labs based on recent research projects from the instructor's dissertation. Example topics include:

• generating text-based explanations for humans making sentencing and bail decisions.
• generating chart-based explanations for sequential decisions in game-playing domains.