The article explores the military standard MIL-STD-882E, which categorizes software control into four levels: the most dangerous is software directly controlling situations that could immediately lead to hazards; less dangerous is software control with delays or reliance on immediate human response; even less dangerous is software that doesn’t directly control but allows time for verification of recommendations; and the safest is software used only for auxiliary purposes. The author points out that with the development of LLM and computer vision technologies, software is being integrated into processes in unprecedented ways, making risk management more critical. This standard provides a practical framework for evaluating the safety and reliability of AI systems, especially applicable to high-risk fields such as autonomous driving and industrial control, emphasizing the importance of understanding software control levels in the AI era to ensure system safety.
