Satya Nadella said on September 13, 2026, that any pursuit of superintelligence should benefit humanity and remain under human control. The Microsoft chairman and CEO paired that principle with deliberate pacing, embedded evaluators, broader participation in governance and greater control for organizations developing or adopting AI.
The position is a policy direction, not a description of a completed oversight regime. Microsoft planned a public consultation for a Code of Conduct covering its first-party MAI models on September 14, but the document’s contents and force are not part of the announcement.
Nadella’s condition for pursuing superintelligence
Nadella’s central argument is straightforward: powerful AI is worth pursuing only when it helps people and remains subject to human control. In Microsoft’s established Responsible AI language, that control is tied to accountability—people should remain responsible for AI systems and oversee how they operate.
That principle matters because it sets a boundary around the usual race-for-capability narrative. Nadella also called for AI’s benefits to spread broadly, so his position is not a demand for an unconditional halt. It is an argument for making alignment—the work of keeping an AI system’s behavior consistent with human goals—a design priority while development continues.
What deliberate pacing and embedded evaluators would mean
“Deliberate pacing” means moving carefully enough to improve alignment rather than treating speed as the only measure of progress. Nadella presented it as a way to get the design of advanced AI right, not as a blanket moratorium on development.
He also welcomed embedded evaluators: people or teams positioned within AI-development environments to assess safety measures as systems are built. That idea could make oversight part of the development loop instead of a final inspection performed after deployment.
The practical details matter enormously. The proposal does not define how independent evaluators would be, what systems they could inspect, who they would report to or whether they could block a release. A label such as “human oversight” is useful only when people have enough information and authority to make a meaningful decision.
A wider AI ecosystem, not control by a handful of entities
Nadella argued that advanced AI governance should involve countries, academic institutions and multiple fields rather than a small group of companies. He also said that open-source and closed models should both be able to thrive, preserving choice across the AI stack.
For businesses, his position includes another form of control: organizations should retain their own knowledge, continuous-learning loops and model weights instead of becoming dependent on a single provider. A learning loop is the process by which a system uses feedback or experience to improve its behavior. Keeping control of that process affects how much an organization can adapt a model—and how easily it can change providers.
The main ideas can be summarized like this:
| Policy dimension | Microsoft’s stated position | Practical implication |
| Human control | Advanced AI should benefit humanity and remain under human control. | People must remain accountable for how AI systems are governed and used. |
| Development pace | Alignment should be improved through deliberate pacing. | Progress should be measured by more than speed or capability gains. |
| Evaluation | Nadella supports embedded evaluators in AI-development environments. | Oversight could become part of development, subject to decisions about access and authority. |
| Governance | Advanced AI should not be controlled by a handful of entities. | Countries, academic institutions and different fields should have a role. |
| Model ecosystem | Open-source and closed models should coexist. | Organizations retain more choice across the AI stack. |
| Enterprise control | Organizations should control their knowledge, learning loops and model weights. | Dependence on one model provider can be reduced. |
Microsoft’s established responsible-AI framework
Microsoft already organizes its Responsible AI framework around six principles: fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability. Its tools and practices include impact assessments, human-AI design guidance, red teaming, evaluation tools and governance tools.
Those principles provide institutional context for Nadella’s statement, but they are not the same thing as the newer evaluator proposal. Microsoft AI’s “Humanist Superintelligence” concept, published on November 6, 2025, similarly describes advanced AI as domain-specific, limited and controllable rather than an unbounded autonomous entity.
That distinction is easy to lose in corporate AI language. A framework can define values and provide tools; a working oversight system also needs clear access rights, decision authority and accountability when something goes wrong.
The MAI Code of Conduct remains the next milestone
Microsoft planned to publish a Code of Conduct for its first-party MAI models for public consultation on September 14, 2026. Nadella described it as part of an approach built around broad access and choice, enterprise control of learning loops and models, and human oversight.
