Broadening AI Access Through Human-Centered Natural Language Interfaces

Abstract

As natural language becomes the default interface for human-AI interaction, questions arise on how to design large language models (LLMs) to safely support a range of human tasks. In this dissertation, I present three lines of work that contribute to our understanding and design of natural language interfaces, taking the perspective that the safe design of LLMs is not only a technical problem, but one that involves the consideration of human factors. First, I investigate the key safety risks of LLMs in their ability to communicate risk and limitations to humans, finding that LLMs struggle both to interpret and generate expressions of certainty correctly. Next, I propose a new evaluation framework to understand the potential harms of human-LM interactions, emphasizing the need to evaluate the behaviors triggered by generated language rather than the language quality itself. Lastly, I explore the ways in which NLP research could support the needs of a broader user audience, advocating for the introduction of tasks and needs previously overlooked in our literature. Together, my work identifies new safety risks, reveals mitigation solutions, proposes a new evaluation framework, and expands our understanding of how LLMs could be used to safely support the needs of a broader user audience.