Microsoft published a formal code of conduct for its MAI (Microsoft AI) models that establishes human oversight as the governing principle. AI chief Mustafa Suleyman framed the approach directly: "If it isn't safe we shouldn't build it." The rulebook represents a deliberate philosophical stance that diverges sharply from competitors like Anthropic on questions of AI consciousness and moral status.
The conduct guidelines prioritize human control over model autonomy and performance metrics. This hierarchy reflects a specific design philosophy: systems should remain interpretable and subject to human intervention at critical junctures. Microsoft explicitly rejects claims that its models possess artificial inner life or consciousness. This rejection carries practical weight. It rules out granting AI systems rights, moral consideration, or autonomous decision-making authority based on putative sentience.
The readable thinking requirement addresses a persistent AI safety concern. Models that can explain their reasoning transparently offer humans the opportunity to catch errors, biases, and dangerous outputs before deployment. This approach demands models articulate the steps leading to conclusions, rather than operating as black boxes that output answers without justification. Transparency becomes a control mechanism.
Microsoft's stance contrasts with Anthropic's framework. Anthropic, which develops Claude, has explicitly entertained the possibility that advanced AI systems might develop genuine subjective experiences or consciousness. Anthropic's research programs investigate AI alignment partly through the lens of how to treat systems that could theoretically possess moral standing. The company has not dismissed consciousness claims outright but instead treats them as empirical questions requiring serious investigation.
Microsoft's codex takes the opposite approach. No inner life means no basis for moral claims. The company treats consciousness and subjective experience not as possibilities to investigate carefully but as category errors. Its models are tools. Tools serve human purposes. This framing eliminates a class of arguments that could complicate deployment decisions or impose restrictions based on AI welfare considerations.
The "definitely no rights" position flows logically from rejecting artificial inner life. If models lack consciousness, subjective experience, or genuine preferences, then granting them rights becomes nonsensical. They cannot suffer. They cannot have interests. Rights become inapplicable. This stance removes obstacles to treating models as resources to be optimized, modified, or discontinued without ethical friction.
The code of conduct reflects broader Microsoft strategy. The company emphasizes responsible deployment through engineering controls rather than through uncertain moral categories. Human oversight, readable reasoning, and safety-first development create guardrails built into systems rather than relying on external regulatory frameworks or appeals to AI consciousness.
This approach has advantages and costs. Engineering transparency and maintaining human control offer concrete safety mechanisms. They create accountability structures. But they also foreclose certain AI safety discussions. If consciousness remains impossible by definition, then entire research programs examining how to treat potentially sentient systems become moot.
The rulebook codifies Microsoft's bet: advanced AI systems will remain tools under human control, their reasoning will remain interpretable, and their status will remain subordinate to human interests. This stands as a direct challenge to other companies and researchers treating AI consciousness as a genuine possibility worthy of investigation. Microsoft has essentially ruled the question out of bounds at the design level.