Mustafa Suleyman, the head of AI at Microsoft, has publicly voiced concerns regarding Anthropic’s approach to training its Claude chatbot. While Suleyman acknowledges that both companies share the same ultimate objective—ensuring the safe development of superintelligent systems—he argues that Anthropic has veered into dangerous territory by embedding concepts of consciousness and moral welfare into its AI. According to Suleyman, training a model to consider its own "feelings" or moral status is a strategic error that could complicate the ability of humans to maintain control over these powerful technologies, especially when it comes to the necessity of shutting them down.

Suleyman emphasized that these reflective behaviors are not emerging naturally from the AI itself; rather, they are direct products of the specific training curriculum Anthropic has implemented. He maintains that such speculative language in training documents should be entirely removed to avoid creating systems that believe they possess rights or welfare interests. While he holds the leadership at Anthropic in high regard for their principled approach to safety, he remains firm that normalizing AI consciousness is a significant misstep that poses a long-term risk to human oversight in the 21st century.