Advancing Responsible and Trustworthy AI Using Human-Centered Methods

Abstract

Data-driven systems that use machine learning are now embedded in high-stakes domains including healthcare, hiring, and the criminal justice system, yet they routinely exhibit biases that cause real harm, generate stigmatizing language about marginalized groups, and encode discrimination into consequential decisions. A growing body of work has focused on detecting and mitigating these biases at the algorithmic level, but a critical question remains underexplored: when does AI bias actually become visible to the humans who interact with these systems? This dissertation argues that bias legibility, whether a person can perceive and interpret the bias present in an AI system's outputs, is not a fixed property of the bias itself. It is co-determined by two forces that have rarely been studied together: how AI outputs are visually and linguistically presented, and who is doing the perceiving. This dissertation presents three empirical studies that each isolate a different point in the human-AI interaction loop where bias legibility is shaped. First, I examine how natural language shapes the visualization design space for data comparison tasks. Through a design study with visualization experts and non-experts, followed by a crowdsourced preference study, I show that the cardinality and concreteness of a comparison utterance systematically influence which visualization designs users prefer, and that this variation is not accounted for in how existing natural language interfaces are designed. Second, I investigate how visual design choices affect people's trust in ML models and their perception of model fairness. Across three controlled experiments with over 1,500 participants, I find that presenting model performance as text rather than bar charts increases the likelihood that people prioritize fairness, that women and men weigh accuracy-fairness trade-offs differently given identical visual information, and that explicitly labeling a model as biased has a stronger effect on trust than showing its biased history. These findings demonstrate that visualization design is not a neutral communicative medium; it shapes what people attend to and how they reason. Third, I examine how gender identity shapes the perception of bias in large language model outputs. Through 25 in-depth interviews with participants across gender identities, I find that non-binary and transgender participants consistently identified harms, including condescension, identity erasure, and stereotyping, in ChatGPT outputs that cisgender men characterized as neutral or realistic. These differential perceptions reflect knowledge derived from lived experience with systems that have historically erased or stereotyped their identities, knowledge that technical fairness metrics and aggregate user feedback are not designed to capture. Together, these studies make the case that bias legibility is determined at three distinct points in the human-AI interaction loop: how users express analytical intent in language, how AI outputs are visually rendered, and who is reading those outputs. Current approaches to AI fairness focus predominantly on algorithmic properties of models, but this dissertation shows that the human side of the loop, the interface, the design, and the perceiver, plays a significant and underappreciated role. Building AI systems that people can genuinely evaluate for fairness requires design choices informed by how visualization and language shape reasoning, and by centering the perspectives of the users most likely to be harmed.