This preprint surveys the role of humans as explicit safety constraints in reinforcement learning (RL) for safety-critical systems. Unlike traditional human-in-the-loop RL approaches that focus on learning efficiency, this work emphasizes human oversight to prevent catastrophic outcomes in domains such as autonomous driving, medical robotics, and industrial control. Using a systematic PRISMA-based review of 100 studies from 2010–2025, the article identifies gaps in purely algorithmic safety approaches and introduces the Human Safety Constraint Framework (HSCF), which formalizes human roles as preventive, corrective, advisory, and normative constraints. Case studies illustrate how human intervention mitigates residual risks, and the survey concludes with recommendations for developing scalable, certifiable hybrid human-algorithm safety architectures.