AI Safety Landscape for Large Language Models: Taxonomy, State-of-the-art, and Future Directions

AI Safety requires a three-part approach: The paper identifies Trustworthy AI, Responsible AI, and Safe AI as interconnected perspectives needed to address safety comprehensively. This means organizations must consider not just whether AI works reliably, but also whether it's deployed responsibly and whether it prevents harm—rather than treating these as separate concerns.

Large Language Models have expanded the scope of AI Safety beyond traditional concerns: The rapid advancement of generative AI has broadened safety considerations to include impacts on "public safety and national security," suggesting that modern AI systems pose risks that go beyond individual user harms to affect society at large.

Current AI Safety requires innovative testing and design mechanisms: The paper emphasizes that new "methodologies and techniques for designing and testing AI safety" are needed specifically for large language models, indicating that existing safety frameworks may be insufficient for these powerful new systems and that the field is still developing best practices.

Building public trust depends on addressing AI Safety systematically: The paper's stated goal to "enhance people's trust in digital transformation" through safety research suggests that without comprehensive attention to safety across multiple dimensions, public adoption and acceptance of AI technologies will remain limited.