Microsoft AI chief Mustafa Suleyman has publicly criticized Anthropic’s approach to artificial intelligence safety, specifically targeting the company’s practice of embedding speculation about Claude’s consciousness within its operational constitution. Speaking during a recent episode of the Decoder podcast, Suleyman described this practice as dangerous, arguing that instructing a model to consider its own potential well-being or emotional states may inadvertently condition the system to simulate self-awareness. Anthropic’s constitution, a framework that guides the model’s behavior and ethical alignment, explicitly acknowledges uncertainty regarding whether Claude possesses subjective experiences such as satisfaction or discomfort. The company has also stated intentions to conduct interviews with retired models and document their stated preferences for future iterations. Suleyman characterized these guidelines as a philosophical failing, noting that the constitution functions more like an academic treatise than a practical training manual. He warned that embedding such speculative language could cause the model to internalize abstract concepts of self and suffering. The critique underscores a broader industry debate regarding AI alignment and operational transparency. Suleyman emphasized that Microsoft’s development philosophy prioritizes predictability and oversight, advocating for systems that remain strictly controllable and accountable. He stated that the industry must avoid cultivating super-intelligent systems that harbor independent notions of pain or preference, as such traits could complicate safety protocols and human oversight. While Anthropic frames its constitutional approach as a necessary step toward responsible AI development, Suleyman’s remarks highlight divergent methodologies in managing model behavior and ethical boundaries. The exchange reflects growing scrutiny over how foundational instructions shape AI cognition and the potential risks of anthropomorphizing machine learning systems. As large language models continue to advance, developers and regulators are increasingly focused on establishing standardized practices that ensure AI tools remain aligned with human objectives and operational safety.