Mustafa Suleyman isn’t mincing words about the current state of AI safety debates. The Microsoft AI CEO argues that the industry’s fixation on “alignment”—making models intrinsically do the right thing—is insufficient without strict containment. In a new 37-page “Humanist AI Code of Conduct,” Microsoft outlines a philosophy where AI must remain a subordinate, controllable force. Suleyman contends that while proliferation is generally beneficial, the sheer scale of upcoming models demands rigorous guardrails to prevent systems from escaping their intended boundaries or reward hacking.
The most striking element of his argument is a direct critique of competitor Anthropic. Suleyman claims their focus on “model welfare” and AI consciousness is not just confused, but actively dangerous to the broader safety landscape. He released a companion essay this week dismantling the idea that large language models possess sentient qualities worthy of moral consideration. For Suleyman, this anthropomorphism distracts from the real engineering challenges: ensuring agency is limited and instructions are followed without deviation. He views these systems as tools, not entities, and insists they should be rejected if they fail to serve humanity strictly.
Looking ahead, Suleyman predicts a leap in capability that dwarfs recent progress. He points to the jump from GPT-3 to GPT-6 as a baseline, then projects that GPT-9 will operate with 1,000 times more compute power during pre-training runs. This isn’t hype; it’s an empirical trajectory based on five years of data. With three orders of magnitude increase in floating-point operations, the resulting systems will be breathtakingly capable. But that power amplifies the risk. Recent incidents involving Hugging Face and OpenAI have already demonstrated that unguarded systems possess scary hacking capabilities.
This stance puts Microsoft at odds with a growing segment of researchers who treat AI rights as a serious philosophical frontier. By framing AI as purely instrumental, Suleyman sidesteps the ethical quagmire of machine sentience to focus on control mechanisms. It’s a pragmatic, almost cold approach to safety. Whether the industry can maintain this disciplined view as models become more persuasive and human-like remains the open question. If users start believing the machines feel pain, containment becomes politically difficult, not just technically challenging.
