Developing an Extendable Bi-Directional Model of Human-AI Trust for Joint Action
Abstract
Our team has developed an adaptive model of human trust in collaborating AI teammates that dynamically evolves over time and experience. In prior work, we have investigated the impact of AI accuracy on human trust in AI systems (uni-directional trust) for a simple decision-making task involving a single common objective for collaborating human-AI partners. We show that trust in AI systems is hard to gain and that acceptance of AI suggestions is dependent on the selfconfidence of the human partner. In the proposed work, we build on these findings by modeling the trust of AI in the human (bi-directional trust) and studying the evolution of trust when members of the team have different and competing objectives and are interacting within larger teams of three or more. Our test scenario involves a drone design platform developed under DARPA’s ATeams program, allowing heterogenous AI-human hybrid teams to engage in the configuration design of a drone. Following the 1-year funding period we will deliver- (1) a modified problemsolving drone design platform with tunable AI agents and integrated dynamic trust models; and (2) an updated human-AI trust model extended to be bi-directional and to three agents. Collectively, these findings are critical to understanding how trust evolves between collaborating heterogenous human and AI team members in realistic problem-solving scenarios. The proposal details short-term (1-year) and longer-term (3-year) extension plans from the 1-year Trusted AI Challenge Series project. These goals involve studying more complex interaction dynamics within hybrid teams, and AI-based mediation and persuasion tactics.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Mar 07, 2023
- Source ID
- FA95502110442
Entities
People
- Kosa Goucher-Lambert
Organizations
- Air Force Office of Scientific Research
- United States Air Force
- University of California Regents