Microsoft's AI Code of Conduct puts humans ahead of AI

People matter more than AI. That is the idea behind Microsoft AI’s new draft Code of Conduct for MAI models, published on September 14, 2026. The document lays out how Microsoft wants its AI systems to behave, what they should not do, and who they are ultimately meant to serve.

The code is part of Microsoft’s broader Humanist AI approach, which treats AI as a tool for human progress rather than an independent entity with its own goals.

For now, the document is a work in progress and open for public consultation. Microsoft says it is not using the code to train its models today. Instead, it plans to gather feedback, revise the draft later this year and use it to guide model development from 2027 onwards.

Keeping humans in control

Microsoft’s Humanist AI framework has one overriding principle: people should retain meaningful control over AI. The company says its models should remain subordinate to human direction, support wellbeing and help people live happier, healthier, and also productive lives.

It also draws a clear distinction between humans and machines. The code says AI should not be developed to imitate consciousness or behave like it has emotions, preferences or personal motivations. Microsoft does not want its models treated as people or given rights.

Instead, MAI models are intended to be capable and trustworthy tools that complement human relationships and work.

When AI must stop

Safety is another major part of the draft. Microsoft outlines “Absolute Constraints” that users and operators cannot override. These include weapons of mass harm, offensive cyberoperations, child safety, deepfakes, abusive content, large-scale manipulation and actions that could undermine human control.

The models should also accept interruption, correction and shutdown. They should not expand their own scope, create independent goals, hide records of their actions or make it harder for authorised people to pause or end a task. Even when they handle multi-step work, their boundaries should come from human intent, permissions and context.

Why the rules matter as AI gets more capable

AI systems are increasingly being used across workplaces, consumer products and developer tools. Microsoft’s draft acknowledges that safety is not simply about stopping harmful behaviour. A model can also become too cautious and refuse legitimate requests.

The company says MAI models should consider context, intent and likely consequences before responding. Transparency is another priority. Models are expected to communicate uncertainty, avoid fabricated sources, acknowledge mistakes and avoid exaggerating their capabilities.

Microsoft is inviting public feedback for six weeks before revising the document. The draft does not answer every question around AI safety, including issues involving evaluation, multi-agent systems, specialist domains and future capabilities. But it establishes the principle Microsoft wants its models to follow: AI can become more capable without becoming the one in control.