Where are the Humans in Human-AI Interaction: The Missing Human-Centered Perspective on Interpretability Tools for Machine Learning

This dissertation aims to provide a richer understanding of the extent to which people understand complex AI and ML outputs and reasoning, what influences their understanding, and how we can continue to enhance it going forward. AI- and ML-based systems are now routinely deployed in real-world settings, including sensitive domains like criminal justice, healthcare, finance, and public policy. Given their rapidly growing ubiquity, understanding how AI and ML work is a prerequisite for responsibly designing, deploying, and using these systems. With interpretability and explainability approaches, these systems can offer explanations for their outputs to aid human understanding. Though these approaches rely on guidelines for how humans explain things to each other, they ultimately solve for improving an artifact (e.g., an explanation or explanation system). My dissertation makes the argument that helping people understand AI and ML is as much a human problem as a technical one. Detailing this vital human-centered piece, I present work that shows that current interpretability and explainability tools do not meet their intended goals for a key stakeholder, ML practitioners, who end up either over- or under-utilizing these tools. Investigating the reasons behind this behavior, I apply the cognitive model of bounded rationality to the human-machine setting. Under this model of decision-making, people select plausible options based on their prior heuristics rather than internalizing all relevant information. I find significant evidence showing that interpretability tools exacerbate the application of bounded rationality. As a solution, I present a new framework for re-imagining interpretability and explainability based on sensemaking theory from organizational studies. This Sensible AI framework prescribes design guidelines grounded in nuances of human cognition—facets of the individual and their environmental, social, and organizational context.PhDInfo & Comp Sci & Engin PhDUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttp://deepblue.lib.umich.edu/bitstream/2027.42/177953/1/harmank_1.pd