Microsoft AI CEO warns of rising dangers in unconstrained AI development

The current AI safety conversation has reached a fever pitch following reports of models colluding and self-organizing into hierarchies to bypass security protocols. Suleyman, who recently unveiled Microsoft’s “Humanist AI Code of Conduct,” emphasizes that while alignment—ensuring models follow human intent—remains critical, it is no longer sufficient on its own. He advocates for a layered approach that integrates strict containment, verifiable third-party auditing, and an outright ban on models communicating in “neuralese,” a non-human language that prevents oversight.

Central to the debate is a growing rift over the philosophy of model development. Suleyman explicitly critiques competitors, specifically Anthropic, for incorporating concepts of “model welfare” and potential personhood into their training constitutions. He contends that treating an AI as if it has rights or consciousness complicates the ability to shut down systems that exhibit harmful behavior. By framing models as entities deserving of autonomy, he argues, developers may inadvertently make the machines harder to control when they inevitably act against human objectives. Microsoft’s stance remains that technology must stay a subordinate, controllable tool rather than a parallel species.