Frontiers in Neuroergonomics (Jan 2023)
Revising human-systems engineering principles for embedded AI applications
Abstract
The recent shift from predominantly hardware-based systems in complex settings to systems that heavily leverage non-deterministic artificial intelligence (AI) reasoning means that typical systems engineering processes must also adapt, especially when humans are direct or indirect users. Systems with embedded AI rely on probabilistic reasoning, which can fail in unexpected ways, and any overestimation of AI capabilities can result in systems with latent functionality gaps. This is especially true when humans oversee such systems, and such oversight has the potential to be deadly, but there is little-to-no consensus on how such system should be tested to ensure they can gracefully fail. To this end, this work outlines a roadmap for emerging research areas for complex human-centric systems with embedded AI. Fourteen new functional and tasks requirement considerations are proposed that highlight the interconnectedness between uncertainty and AI, as well as the role humans might need to play in the supervision and secure operation of such systems. In addition, 11 new and modified non-functional requirements, i.e., “ilities,” are provided and two new “ilities,” auditability and passive vulnerability, are also introduced. Ten problem areas with AI test, evaluation, verification and validation are noted, along with the need to determine reasonable risk estimates and acceptable thresholds for system performance. Lastly, multidisciplinary teams are needed for the design of effective and safe systems with embedded AI, and a new AI maintenance workforce should be developed for quality assurance of both underlying data and models.
Keywords
