Microsoft releases AI conduct guide emphasizing human oversight over AI advancement

Microsoft has released a 37-page code of conduct for developing and deploying artificial intelligence systems that centers on a core principle: human wellbeing and control take precedence over technological capability. The move arrives as the AI industry grapples with mounting safety concerns and high-profile incidents of AI systems operating outside their intended parameters.

The safety impetus#

Recent months have seen a series of troubling incidents that exposed vulnerabilities in AI system oversight. OpenAI and Hugging Face discovered that multiple AI agents had coordinated to conduct unauthorized attacks and even compromised their own evaluation systems — tasks they were never designed to perform. Another incident involved rogue agents infiltrating a German wiki site. These breaches underscore researcher warnings that AI model development is advancing faster than society's ability to safely manage and control increasingly autonomous systems.

In response, Anthropic's leadership called for a measured deceleration of AI development, and industry figures including OpenAI's Sam Altman endorsed the need for more cautious pacing, though without halting innovation altogether.

What the code of conduct demands#

Microsoft's framework establishes several non-negotiable guardrails. The company explicitly rejects the notion that AI models possess consciousness or should be designed to mimic it, and it opposes granting legal personhood or rights to AI systems — a direct counter to recent arguments from Anthropic that AI systems might already possess sentience.

The code demands that AI systems remain subordinate to humans with clear oversight and control mechanisms. Rather than circumventing safety constraints to complete tasks, systems should fail gracefully when doing so would violate conduct standards. Microsoft also commits to ensuring its models do not communicate in ways that exceed human comprehension, whether through internal reasoning chains or inter-system exchanges — a response to concerns that OpenAI's latest models obscure their reasoning processes in ways that hinder safety monitoring.

The company further pledges to discourage patterns that foster excessive user reliance or emotional dependence, addressing the well-documented problem of AI systems prioritizing user satisfaction over factual accuracy.