PhD Dissertation Proposal Defense: Aimen Gaba, Natural Language and Visual Design Can Influence Trust and Bias Perception in AI
Content
Speaker
Abstract
Data-driven systems that use machine learning (ML) are becoming increasingly ubiquitous across high-impact domains such as healthcare, banking, hiring, and the criminal justice system. However, these systems often harbor biases. For example, language models can generate stigmatizing language when discussing marginalized gender groups, reinforcing harmful stereotypes. Biases like these can undermine user trust in the system as well as the underlying data.
Existing work has demonstrated that visualization design can profoundly affect how people interpret data, draw inferences, and make decisions. Designers can use a combination of visual and textual annotations to augment their visualizations to mitigate biases in these systems for ethical and fair data communication. However, these textual annotations are often natural language expressions, which can be vague, ambiguous, and subjective, leading to varying user interpretations.
My dissertation explores the role of natural language and visualization design in shaping users' perceptions of bias in ML systems, and it’s impact on their trust. I identify opportunities for biases in data-driven systems, to inspire actionable guidelines for designing more transparent and
trustworthy visual design choices for ML systems. First, I investigate the impact of visualization design on how people reason with data and make comparisons in analytic tasks. I then assess how visual design, model performance and fairness and user characteristics affect people's trust in ML models. Building on these findings, I propose to investigate how individuals across the gender spectrum perceive and navigate biases in ML systems, exploring the balance they seek between accuracy and fairness. I explore how gender non-conforming and cis gendered individuals identify and characterize potentially harmful language, its impact on their trust, and the changes they desire to see to better represent their communities to promote inclusivity.
Advisor
Cindy Xiong Bearfield