Model-Based Systems Engineering (MBSE) has become a mandated practice for Department of Defense acquisition programs, yet measured benefits remain elusive. The 2024 Defense Science Board found that less than one percent of published literature actually quantified MBSE outcomes, and flagship ground vehicle programs such as the XM30 Infantry Fighting Vehicle have experienced schedule delays attributed directly to insufficient proficiency with model-based approaches. This paper presents the Digital Safety Twin concept: an AI-powered safety intelligence architecture that addresses three of the most labor-intensive and error-prone MBSE workflows. First, the architecture uses hybrid natural language processing and large language model (NLP/LLM) pipelines to auto-formalize unstructured natural language documents into formally structured, traceable requirements. Second, it auto-generates and continuously maintains traceability relationships across requirements, design elements, hazard analyses, and verification artifacts. Third, it provides continuous safety case completeness and confidence assessment through automated Goal Structuring Notation (GSN) synthesis connected to live evidence sources. The approach is grounded in Systems-Theoretic Process Analysis (STPA), the OMG Risk Analysis and Assessment Modeling Language (RAAML), MIL-STD-882E system safety practice, and the UL 4600 safety case framework. We present the methodology, its alignment to the DoD Digital Engineering Strategy, and its applicability to ground vehicle autonomy programs including next-generation infantry fighting vehicles and robotic combat vehicles. We also discuss the limitations, risks, and cultural barriers that must be addressed for AI-augmented safety engineering to achieve acceptance in mission-critical defense applications.