Microsoft is drawing a hard line in the sand for its next generation of artificial intelligence. The tech giant released a new code of conduct for its MAI models that prioritizes human oversight above all else, even if it means sacrificing performance or autonomy. Mustafa Suleyman, head of Microsoft AI, put it bluntly to The Information: “If it isn’t safe we shouldn’t build it.” This isn’t just marketing fluff. The company is willing to throttle generality to keep humans in the loop. The code will undergo public consultation now, with a final version ready by late 2026 to guide development starting in 2027.
One of the most striking mandates is the ban on “Neuralese.” Microsoft insists its models must not use internal communication methods that humans cannot understand. If people can’t read the reasoning trace, they can’t oversee the action. This stance directly challenges recent trends where models like OpenAI’s GPT-6 Astra have made reasoning traces harder to monitor, even as they become safer. Microsoft acknowledges that readable thoughts aren’t a silver bullet against manipulation, but refusing opaque internal languages is a concrete step toward accountability. Subagents face the same restrictions; they cannot expand their scope or hide actions without fresh approval.
The real divergence appears when comparing Microsoft’s approach to Anthropic’s. While both companies use top-level behavioral documents, their philosophical foundations differ sharply. Anthropic treats Claude as a novel entity with potential moral status, exploring “functional emotions” to improve safety. Microsoft rejects this entirely. Suleyman has long argued against anthropomorphizing AI, stating that agents should have no more rights than a laptop. Microsoft’s models are forbidden from mimicking consciousness or claiming an inner life. This isn’t just about safety; it’s a deliberate effort to strip away the illusion of sentience that plagues current user interactions.
This move aligns Microsoft with CEOs from Meta, xAI, and OpenAI who have recently called for a slowdown in AI development. Microsoft is even open to external auditors verifying its pace. Yet, the 2027 timeline reveals a gap between rhetoric and immediate implementation. The code doesn’t currently train the models, and it doesn’t automatically apply to third-party apps running on Azure. As the industry races toward agentic systems, Microsoft’s bet is that clarity beats complexity. Whether users prefer a transparent tool over a mysterious companion remains to be seen.
