CS547 Human-Computer Interaction Seminar (Seminar on People, Computers, and Design)
Fridays 11:30am-12:30pm PT · Gates B3 · Open to the public|
Qian Yang
Cornell University
Designing Interactions and Anticipating Harm in the Era of Pre-Trained Language Models: An Interaction Designer's Perspective
January 10, 2025
The widespread adoption of pre-trained Language Models (LMs) seems to have significantly simplified the process of prototyping novel, natural language interactions. But is this truly the case? This talk argues that it is not, and that the real challenge of prototyping in the age of LMs lies in evaluation. HCI's prototyping and evaluation methods have evolved over the last several decades primarily under the dominance of the graphical user interface and, therefore, are not directly applicable to language interactions. Moreover, altering an LM application's utility can be as simple as adding a malicious prompt, and this raises new questions about how to set guardrails around them to prevent misuse and other unintended consequences. In this talk, Yang presents her lab's research on designing GPT-powered writing assistants as case studies and explores novel methods to evaluate the human impact of LM-powered systems. This research addresses the following questions: When developing novel GPT-powered autocompletion systems, how can designers assess each design's risk of causing cognitive automation vis-a-vis its benefits to writer creativity? As these systems enter the real world, how can designers not only anticipate potential unintended consequences but also establish guardrails to prevent misuse and appropriation? |
|