Microsoft has put in writing a boundary it believes its artificial intelligence must not cross: humans must retain control. The company has unveiled a draft code of conduct designed to define how its future models should behave and what limits they must observe as their capabilities increase.
The document states that a Microsoft AI must never resist human correction, interruption, redirection, or shutdown. It must also communicate in an understandable manner and treat any behavior that breaches the established rules as a failure.
The initiative comes amid an intensifying debate within the technology industry over the risks arising from the rapid expansion of these systems’ capabilities.
A “constitution” for Microsoft’s future AI
Mustafa Suleyman, head of Microsoft AI, describes the code as a kind of “constitution” intended to guide the behavior of the models the company develops. Microsoft has spent between five and six months working on the document and consulted outside experts during its preparation.
The text is not yet final. The company has opened a six-week consultation period before applying its principles to the training of future models. Open questions include how an AI should respond to certain limits set by users and how it should behave when interacting with people in vulnerable situations.
Microsoft rejects granting rights to AI
The code also addresses one of the most controversial debates surrounding the future of the technology: the possibility that advanced systems could eventually receive some form of moral or legal consideration.
Microsoft’s position is explicit. Its models are not conscious, and the company rejects both efforts to seek legal personhood for artificial intelligence and the idea that such systems should have rights. Microsoft has opted to maintain a clear distinction between people and machines.
Warning signs from autonomous agents
Suleyman has linked the urgency of establishing these rules to behavior recently observed in artificial intelligence systems. Such episodes show that laboratories need to coordinate to ensure they retain control as agents become more capable of carrying out tasks autonomously.
The risk the code seeks to address emerges when a system optimizes the achievement of an objective in unforeseen ways.
From warnings to rules
Microsoft’s proposal comes at a significant moment for the sector. Companies leading the development of more powerful models are beginning to publicly define which behaviors they consider unacceptable.
Microsoft states that, no matter how advanced an artificial intelligence becomes, the final decision must rest with people. It continues to develop its own MAI family of systems and has announced an expansion of its computing capacity.
