Trustworthy, Responsible, and Safe AI: A Comprehensive Architectural Framework for AI Safety with Challenges and Mitigations
Chen Chen, Xueluan Gong, Ziyao Liu, Weifeng Jiang, Si Qi Goh, Kwok-Yan Lam
TL;DR
The paper proposes a holistic architectural framework for AI safety built on three pillars—Trustworthy AI, Responsible AI, and Safe AI—to address safety in the Generative AI era. It surveys foundation-model concepts, lifecycle stages, and formal definitions of safety, then details challenges across inputs, adversarial threats, and ecosystem-level risks, followed by cross-cutting mitigation strategies (red teaming, safety training, guardrails, decoding, capability control, alignment, and governance). Key contributions include a structured taxonomy of risks (from jailbreaking to data privacy and multi-agent threats) and a comprehensive set of mitigation approaches, including Recursively Refined Reward Modeling and cross-distribution interventions. The work emphasizes ecosystem-level safety, governance, and future directions such as comprehensive evaluation frameworks, domain knowledge integration, and defensive AI systems, aiming to enhance public trust and safe digital transformation in complex AI ecosystems.
Abstract
AI Safety is an emerging area of critical importance to the safe adoption and deployment of AI systems. With the rapid proliferation of AI and especially with the recent advancement of Generative AI (or GAI), the technology ecosystem behind the design, development, adoption, and deployment of AI systems has drastically changed, broadening the scope of AI Safety to address impacts on public safety and national security. In this paper, we propose a novel architectural framework for understanding and analyzing AI Safety; defining its characteristics from three perspectives: Trustworthy AI, Responsible AI, and Safe AI. We provide an extensive review of current research and advancements in AI safety from these perspectives, highlighting their key challenges and mitigation approaches. Through examples from state-of-the-art technologies, particularly Large Language Models (LLMs), we present innovative mechanism, methodologies, and techniques for designing and testing AI safety. Our goal is to promote advancement in AI safety research, and ultimately enhance people's trust in digital transformation.