Reinforcement learning has revolutionized how we train AI, enabling it to learn complex tasks through trial and error. However, ensuring these AI agents remain safe and aligned with human values presents a significant challenge. Traditional approaches, which treat safety as a mere constraint in the learning process, can lead to unforeseen consequences, with AI prioritizing overall performance over consistent safety. Think of it like a self-driving car optimizing for speed while only *averaging* its adherence to traffic laws – it might average out okay but still run red lights occasionally. This “safety interference” phenomenon, where prioritizing one safety aspect compromises others, highlights the limitations of current methods. A new approach called Rectified Policy Optimization (RePO) aims to address this issue. Instead of averaging safety across all scenarios, RePO penalizes *any* safety violation during training. This method focuses the AI's learning on truly safe behavior, rather than allowing it to trade off safety in certain situations for better overall performance. Imagine that self-driving car now being penalized for *every* traffic infraction – it would learn to be consistently safe, not just on average. Experiments with RePO on language models have shown promising results, demonstrating enhanced safety across diverse prompts while maintaining performance. RePO shows a path towards more reliable and trustworthy AI, highlighting the importance of integrating human feedback deeply into the learning process.