Presentation
Concept Cards for Human-Centered AI
SessionLBR3: Late-Breaking Results
DescriptionThere are many types of “cards” for developing Artificial Intelligence (AI) such as model cards, data cards, and test cards, which serve as concise and structured representations of some aspect of an AI-based technology. However, most of these cards are mechanisms for AI developers to communicate their work to others—often other AI developers. Moreover, the existing cards all assume the right AI is being built. It is necessary to create tools that help users communicate their needs to AI developers, resulting in AI that is tailored to the problem at hand. We propose “Concept Cards” as a means to represent and convey user needs for decision support to AI developers. The Concept Cards tool employs a two-phase process to both elicit and represent user-centered concepts for AI, so they can be effectively communicated to AI developers in useful manner. We conducted a pilot study (N = 14) using four example concept cards where participants with AI, human factors, and domain expertise all agreed that Concept Cards provided utility across the systems engineering lifecycle (i.e., supported requirements writing, design activities, test and evaluation). Further, participants reported that Concept Cards were enjoyable to use, and forced them to think about the AI and user needs in ways they normally would not have, making Concept Cards valuable as a tool for developing more human-centered and trustworthy AI. Here we will present the Concept Card tool, results from the pilot study, and directions for future work based on these results.
Event Type
Late Breaking Results
TimeThursday, September 12th3:40pm - 3:50pm MST
LocationFLW Salon D