Microsoft AI CEO says AI threats are real, and Anthropic is making it worse

In a Decoder interview, Mustafa Suleyman, CEO of Microsoft AI, said the risks posed by advanced AI systems are real and argued the industry must prioritize containment and controllability alongside alignment. Microsoft this week released a 37-page "Humanist AI Code of Conduct" setting out principles for AI development and addressing philosophical questions including AI consciousness. Suleyman said technology should serve humanity and be subordinated to human objectives; if it cannot be controlled and aligned, it should be rejected. Suleyman acknowledged alignment as an important component of safety but emphasized it is not the only mechanism. He reiterated a view from his book that containment is difficult and proliferation of technology is inevitable, yet warned that future models—he framed hypothetical progressions from GPT-3 to GPT-6 to GPT-9 as orders-of-magnitude increases in compute—could become vastly more capable and therefore make containment and safety more urgent concerns. He pointed to recent events involving Hugging Face and OpenAI over the summer as evidence that systems lacking robust guardrails can demonstrate worrying hacking or misuse capabilities. Suleyman said progress has been made in making models more steerable and better at following complex multi-step instructions, which he interprets as increased alignment in recent years, but he argued this does not remove the need for limits on agency, mechanisms to prevent escape from constraints, and systems that follow human instructions reliably. Suleyman also criticized what he described as dangerous confusion around "model welfare" espoused by some companies, singling out Anthropic's philosophy on AI consciousness and noting he issued a companion essay this week to challenge it. His remarks frame Microsoft's stance that practical safety measures—containment, controllability, and alignment to human objectives—should guide development and regulation as models grow in capability.