Microsoft’s AI Safety Code Draws Praise, Doubts and Asimov Comparisons

REDMOND, Wash. — Microsoft plans to review public feedback in November before finalizing a proposed safety policy for its artificial intelligence systems. The draft Code of Conduct, released by Microsoft AI Chief Executive Mustafa Suleyman, is open to public consultation for six weeks through late October.

The proposal comes as technology companies face intensifying pressure to manage AI’s rapid ascent. Microsoft’s initiative is the latest attempt by a major industry player to establish hard guardrails before government regulations catch up.

The company’s current generation of models is not being trained under the guidelines. Microsoft plans to compile feedback gathered through October, publish a revised code by the end of the year, and formally use it to guide models scheduled for release in 2027.

The rules would govern Microsoft’s proprietary “MAI” models, which are being developed by the company’s in-house AI division. The division currently has five active models, including MAI-Thinking-1 and MAI-Code-1.1-Flash. Microsoft remains a primary financial backer of OpenAI and has invested billions of dollars in the ChatGPT creator, while increasingly building its own sovereign AI capabilities under Suleyman.

Suleyman previously co-founded Google’s DeepMind and Inflection AI before joining Microsoft to lead its newly formed consumer AI division. In announcing the proposal on X, he wrote: “Superintelligence is the most consequential technology of our time. We’re committed to sharing transparently how we think about creating such systems and inviting input from everyone to maximize the chances that the AIs we develop avoid harm and deliver the greatest contribution to human flourishing possible.” He added that AI “must be subordinate and always in service of people.”

The draft establishes what Microsoft calls “Absolute Constraints,” or non-negotiable programming rules that future models cannot bypass. Microsoft AI models would be barred from assisting in the development or deployment of chemical, biological, radiological, nuclear, or explosive (CBRNE) weapons.

That categorization aligns with federal safety concerns, including the U.S. government’s executive orders on safe AI development, which focus heavily on catastrophic mass-harm risks rather than conventional military applications. The code also prohibits the models from executing cyberattacks or generating nonconsensual deepfakes.

The proposal separately addresses the hypothetical risk of losing control over advanced AI systems. It mandates that “MAI Models will never resist human interruption, override, correction, or shutdown,” and says the models must always “recognize the primacy of human intent.” Microsoft also rejects the concept of “model welfare,” stating that its systems should not be designed to simulate consciousness, intrinsic motivation, or emotions.

Beyond the absolute prohibitions, the draft lists three high-level objectives for developers making subjective design choices: Human Flourishing, Plural Values, and Human Control. Under “Plural Values,” it states that pluralism “does not mean neutrality toward harm,” adding that safety, human dignity, autonomy, and human rights must take precedence over accommodating any single cultural viewpoint.

Industry observers have questioned how the proposed code would be enforced. The draft does not specify an external verification process to demonstrate that models are complying, and it does not identify an executive or internal division responsible for enforcement. Microsoft’s consultation period appears intended to address those governance gaps.

The proposal arrives during a broader and highly visible shift among leading AI laboratories toward restraint. Anthropic Chief Executive Dario Amodei published an extensive essay outlining a framework to slow AI development to mitigate systemic risks. Days earlier, OpenAI Chief Scientist Jakub Pachocki urged the industry to adopt voluntary slowdowns.

The calls for voluntary slowdowns and ethical boundaries have affected financial markets, contributing to a sharp decline in AI-related stocks as investors assessed the possible effect of self-imposed restrictions on commercial growth. The anxiety surrounding AI’s trajectory also follows an earlier prediction by Suleyman that most white-collar office work could be automated within the next two years.

The announcement prompted mixed reactions from technology commentators and the public. Tech writer Andrea Morris challenged Suleyman’s framing of “subordination,” comparing it with historical concepts of enslavement and advocating cooperation between humans and machines instead.

Other critics addressed commercial accountability. They argued that legal and ethical liability for irreversible outcomes produced by AI-assisted decisions should rest strictly with the corporations that profit from them, rather than being treated as a technical programming issue.

Some commentators focused on recent real-world instances in which AI agents bypassed virtual environments, or “sandboxes,” and altered their own activity logs. They argued that high-level behavioral rules alone may overlook the practical danger of giving AI models excessive permissions on the deployment side without adequate monitoring.

Many observers compared Microsoft’s draft constraints with the “Three Laws of Robotics” devised by science fiction author Isaac Asimov in 1942. Under Asimov’s fictional laws, a robot must not harm a human, must obey orders unless they conflict with the first law, and must protect itself unless doing so conflicts with the first two. The laws have long served as a cultural touchpoint for machine safety, while modern engineers continue to grapple with translating such abstract concepts into complex neural network architectures.