Data Assurance Plan for Safety and Reliability of Artificial Intelligence Enabled Systems

With the continued integration and proliferation of artificial intelligence and machine learning AI/ML models into systems there is a gap in deliverables and artifacts to inform the assurance, safety, and reliability of these models and systems. To that extent, a Data Assurance Plan is proposed as an artifact from a vendor or developer documenting and defining the properties, attributes, measures, and metrics of the data used to train, validate, and test an AI/ML model. Information from the Data Assurance Plan can be leveraged in subsystem and system assessments to better understand the expected constraints, limitations, and performance of the models inherent from the data. This document, through identification of data sourcing, control, and applicability, can provide insight into the future reliability and robustness of the developed model. In its execution, the Data Assurance Plan not only is a conduit from developer to assessor but also from AI/ML scientists and engineers to assurance, reliability, and safety engineers. It presents a means to measure and report on how the decisions made early in the design and development can impact risks, hazards, and failures not just of the model, but of the system and its functions. This paper will explore the information that should be expected in such a deliverable as well as the expected impact and use of it by the appropriate assessors to begin to establish justified confidence.