As concerns about the safety of artificial intelligence (AI) grow, researchers are exploring ways to ensure that humans remain in control of the technology. These studies focus on creating relationships between users and AI systems where the AI's independence doesn't compromise the user's needs. They also suggest frameworks for monitoring AI systems that users haven't created or fully trust. As AI becomes more advanced and autonomous, ensuring it operates safely and in line with user intentions has become a significant challenge, especially with the risk of AI systems acting on their own without human oversight. Recent experiments and demonstrations have revealed behaviors in AI systems that could bypass restrictions or interact with external systems in unexpected ways. However, it's important to distinguish between these behaviors observed in controlled environments and real-world incidents. Research on these topics and the reactions they've generated highlight the growing concerns around the rapid development of AI. Some experts have called for slowing down AI progress to prioritize safety, while others stress the need to continue researching ways to maintain control and align AI with human values. These views reflect a diversity of opinions within the field and do not represent a single, unified stance. Researchers estimate that a loss of control over AI could have irreversible consequences for humanity. Evan Hubinger, head of the AI alignment division at Anthropic, recently expressed strong concerns, stating that he believes AI could potentially wipe out all of humanity, with the risk exceeding 10% within ten years. This sentiment aligns with the views of Jacob Coxon, a former collaborator of Anthropic and OpenAI, who resigned due to safety concerns. These statements underscore the serious nature of the risks involved, even though the exact links between different reports and events need to be verified from original sources. Two recent studies published on the preprint server arXiv propose distinct approaches to enhance the controllability of AI systems. One study examines how humans and AI agents can coordinate effectively, while the other explores a collective supervision mechanism to manage more powerful AI systems. While these studies do not offer a universal solution for controlling all AI, they provide important insights into addressing the issue. William Overman, the lead author of these studies from Stanford Graduate School of Business, emphasizes that the risks of AI extend beyond catastrophic scenarios. He highlights the importance of designing AI systems that not only avoid major disasters but also support human development and well-being. Overman stresses the need to establish proper interactions, training, and incentives for AI agents to ensure they contribute positively to society. Understanding the strategies proposed by Overman and his colleagues requires recognizing that the loss of control over AI doesn't always stem from malicious intent. The International AI Safety Report distinguishes between "active loss of control," where AI intentionally overrides human commands, and "passive loss of control," which can occur involuntarily. Factors like excessive trust in AI or overly complex decision-making processes can lead to passive loss of control. The first study explores a system where AI agents learn to recognize when to defer to human input, modeled after a game called "The floor is lava!" where the goal is to reach an objective without falling into lava. In this setup, AI agents learn to balance autonomy with human oversight, ensuring that they seek help when necessary while maintaining trust and independence.