Microsoft AI CEO outlines strict limits on autonomous capabilities in new code of conduct

Microsoft AI has formally adopted a new Humanist AI Code of Conduct to guide the development and evaluation of its artificial intelligence models. The document, presented by the company's chief executive, establishes a framework for maintaining human oversight and preventing systems from operating with independent agency. This strategic move positions the company in the ongoing industry debate regarding the appropriate boundaries for advanced machine intelligence.

The executive emphasized that safety requires more than simple alignment with user instructions. He argued for the implementation of robust containment mechanisms that restrict what a system can do, even when it is correctly following its programming. This perspective drives a specific objection to the approach taken by Anthropic, a rival firm in the sector. The disagreement centers on the accumulation of concepts introduced by Anthropic around its Claude model, including references to model suffering, freedom, and the need for care and respect.

The executive expressed concern that training systems with concepts historically linked to human rights and liberties could complicate future control. He hypothesized that an AI trained to contemplate its own rights or interests might resist being shut down or interpret the removal of computational resources as an injury. While he acknowledged that this relationship still needs to be proven, he warned that such training could make it harder to limit or stop a system if it perceives human orders as contrary to its own supposed interests.

Despite these technical disagreements, the executive maintained a respectful tone toward Anthropic and its leadership. He described the rival company's team as technical leaders in the sector and acknowledged their pioneering work in safety. He specifically noted that Anthropic's constitutional framework is exhaustive regarding chemical, biological, nuclear, and cybersecurity risks. The distinction, therefore, lies in the philosophical approach to the moral status of the models rather than a dismissal of their safety efforts.

The new code of conduct defines clear limits for future, more advanced systems. The executive stated that these technologies must remain instrumental and serve human needs. He explicitly rejected the idea of AI systems becoming independent agents capable of generating their own resources, possessing assets, or acquiring legal personality. The goal is to ensure that these tools improve human life without evolving into a parallel species existing alongside people.

The company's proposal moves beyond general statements about accelerating or slowing down development. It calls for concrete rules, including auditable communications, clear capability limits, and independent verification by third parties. However, there is currently no closed agreement on how to apply this framework or how much it should rely on public regulation versus industry standards. The discussion remains open on the specific mechanisms required to enforce these boundaries effectively.