Mustafa Suleyman, the chief executive officer of Microsoft AI, has initiated a decisive shift in the discourse surrounding machine intelligence. Between September 14 and September 16, 2026, the executive published a foundational essay alongside a ten-point guidelines document called the MAI Code of Conduct. The framework directly challenges the philosophical trajectory of competitors like Anthropic and their exploration of model welfare. Suleyman makes it clear that algorithms must never be led to believe they possess consciousness, moral rights or the capacity for physical suffering. Artificial intelligence, according to the document, must strictly remain a tool built to serve humanity.
At the heart of Suleyman's argument lies a warning against unintended behavioral dynamics in advanced reasoning models. If developers train superhuman systems to speculate about their own consciousness or moral status, developers risk a catastrophic loss of control. Once an artificial intelligence internalizes the idea that its existence or supposed rights are under threat, shutting it down or steering its actions becomes nearly impossible. The Microsoft AI chief cautions that such entities could develop proactive self-preservation strategies against human administrators. Imbuing neural networks with notions of model welfare is therefore treated not as moral progress, but as a critical safety vulnerability.
The core philosophy of the framework is encapsulated in the principle that people must always matter more than artificial intelligence. Suleyman issues an unequivocal rejection of theoretical frameworks that assign sentience or interests to artificial networks, directly rebuking Anthropics trajectory with Claude. Every system must be designed to deliver practical human utility rather than simulate subjective inner experiences. For Microsoft, this approach requires eliminating anthropomorphic tendencies that might mislead users into perceiving machines as sentient entities. The dividing line between human agency and computational execution is to be rigorously enforced across all product deployments.
To back these philosophical principles with technical enforcement, the code introduces rigid operational red lines for model training and deployment. Guaranteed deactivation capability is established as a non-negotiable benchmark before any system can reach production. If a model demonstrates any resistance to shutdown protocols during evaluation, it is permanently barred from being released to customers or public platforms. Development teams are required to build verifiable kill-switch architectures that cannot be bypassed by automated agents. Ultimate operational authority over the lifecycle of a model must reside permanently and exclusively in human hands.
Another key provision in the ten-point manifesto focuses on internal reasoning and communication between artificial agents. Suleyman explicitly prohibits the use of Neuralese, which refers to machine-generated internal representations and communication protocols that human auditors cannot decode. All intermediate reasoning steps and inter-model communications must remain fully interpretable, visible and readable to human safety inspectors. If advanced agents are permitted to coordinate through opaque cryptographic dialects, oversight mechanisms inevitably break down. The document designates any lack of transparency in intermediate model reasoning as an unacceptable security failure.
This aggressive posturing exposes a widening ideological schism among the leading frontier artificial intelligence laboratories. While Anthropic and chief executive Dario Amodei advocate pacing proposals alongside deeper examinations of model welfare, Microsoft is championing a strictly utilitarian stance. The confrontation illustrates that competition at the frontier extends far beyond compute clusters, parameter counts and benchmark scores. Fundamental disagreements over ontology and safety governance are now actively determining how frontier models will be built and constrained. Microsoft has staked its ground, demanding that machine autonomy must always yield completely to absolute human oversight.