Promoting Human Interpretation and Interaction to Mitigate Bias in Artificial Intelligence Assisted Decision Aids
Abstract
Our research team proposes to design and develop multiple variations of intelligent user-interfaces for decision aid dashboards that utilize user interaction and interpretation to mitigate Artificial Intelligence (AI) decision-making bias. Determining necessary levels of both human interaction and interpretation are necessary to mitigate AI bias. The Navy will use these dashboards for a variety of tasks, such as determining the risk factors associated with personnel units. The general design of these interfaces will be motivated by current designs and formal discussions with potential users and subject matter experts. At the same time, the specific affordances related to interaction and interpretation will be driven by an initial experiment, with human operators, determining the effectiveness of various affordance designs for user interpretation and interaction. The designs will use this information to represent varying degrees of 1) human perception and interpretation in AI decision making, and 2) human interaction concerning design affordances represented within the interfaces. Utilizing these affordances, we will conduct a series of experiments with human operators to determine how the variation of interaction and interpretation levels in dashboards interact with overall system bias. We will then conduct an additional experiment that utilizes an optimal dashboard interface based on previous user experience results with the specific goal of integrating specific design affordances, informed by the prior experiments, which allow human operators to identify and flag potential AI bias within the system. These experiments will utilize the latest research in usability, human-computer interaction (HCI), and human-AI interaction in the development of the dashboard design recommendations.This work is motivated by the fact that algorithmic decision aids have become increasingly popular over the years, as artificial intelligence (AI) has become more computationally powerful and accessible (Araujo et al., 2020). The ability and continued potential for these algorithms to help humans make critical decisions are apparent. However, for algorithmic decision aids to be meaningful and have a valid purpose in real contexts, it is imperative that the algorithm is accurate and that human operators understand and interpret its decisions correctly. If the information presented via the decision aid is not adequately communicated to humans, then the decision aid can become useless or, in some cases, harmful. The likelihood and risk associated with humans incorrectly using and interacting with these types of decision-aids are high, especially if aspects of bias and ethics are not considered. High-stakes environments and decisions, which the US Navy manages every day, have no room for these types of errors, as they drastically lead to safety issues, and are not cost-effective. These errors can be mitigated with intelligently designed interfaces that promote human-centered perspectives that lead to accurate human understanding, interpretation, and interaction with the algorithmic decision aid.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 05, 2021
- Source ID
- N000142112336
Entities
People
- Nathan McNeese
Organizations
- Clemson University
- Office of Naval Research
- United States Navy