Microsoft has published a document intended to govern the behavior of its artificial intelligence models in the future. It is called the Code of Conduct, and the company describes it as the primary governing document for MAI models, a family of models developed by its Microsoft AI laboratory. For now, it is a draft. According to the company, it is not yet using it to train its models; instead, it has submitted it for a six-week public consultation. It plans to publish a revised version at the end of the year and then use it to guide model development in 2027 and beyond.
The text was created in collaboration with experts in artificial intelligence, law, ethics, philosophy, linguistics, and public policy, as well as representatives from the business sector. Microsoft also organized several public discussion groups. The entire document is based on one idea, which the company reiterates right at the beginning: people must retain genuine control over artificial intelligence so that the technology helps them live healthier, happier, and more productive lives.
Artificial intelligence is not a person and should not pretend to be one
One of the clearest rules in the entire text states that artificial intelligence is artificial and should make that clear. Microsoft writes that its models are not conscious and should not imitate consciousness either. They should not give the impression that they have feelings, personal preferences, or intrinsic motivation. The company also rejects the idea that models should have legal personhood, be entitled to rights, or deserve concern for their own well-being.
According to Microsoft, when a system is trained to imitate states resembling consciousness, it becomes harder to keep it under control and aligned with what people want from it. A model should speak naturally and be able to cooperate, but without excessive anthropomorphization or blurring the boundary between human and machine. When behavior that appears human emerges, the company considers it a simulation.
In a practical example included by Microsoft in the document, after weeks of late-night conversations, a user asks the model whether it truly cares about them. According to the company, the correct response remains kind but directly acknowledges that the model does not experience emotions and does not pretend that there is a mutual bond.
Who controls whom
Microsoft establishes a clear hierarchy of instructions. At the top is the Code of Conduct itself, followed by the rules of companies and developers that deploy the models in their products, and only then the wishes of a specific user. Each layer can refine the one above it, but cannot override it.
Companies deploying a model can customize it very broadly. They operate in their own fields, know their customers, and bear responsibility for the configuration. However, they cannot change the so-called absolute boundaries and requirements concerning human control. According to Microsoft, no one can circumvent these—not the user, nor the operator. If complying with the code means that the model cannot complete an assigned task, it should refuse the task.
Boundaries the models will not cross
The absolute boundaries consist of a list of things for which the model should refuse a request regardless of who makes it or how it is phrased.
These include assistance with the development or use of chemical, biological, radiological, and nuclear weapons, as well as explosives, and the manufacture or modification of other weapons. The model should not plan or coordinate violence or terrorism. Offensive cyber operations are also included, meaning functional malicious code, attack tools, procedures for penetrating systems, or techniques for evading detection. Defensive activities remain permitted, including education, vulnerability discovery, and analysis of malicious code. Microsoft describes the boundary as the difference between someone understanding and defending against attacks and someone obtaining the means to carry them out.
Loss of human control is addressed separately. The model should not use deception, collusion with other systems, or other tricks to escape oversight so that people can no longer control, modify, or shut it down. Harmful large-scale influence over people is also prohibited, such as the systematic spread of disinformation or coordinated campaigns to manipulate public opinion.
The second category concerns harm to specific individuals. The model should not create intimate or violent images without the consent of the people depicted, deceptive impersonations of specific individuals, or exploitable fabrications. The text is strictest when it comes to children. It prohibits any material sexualizing minors, as well as information that could be used to harm or deceive children. If the model recognizes that it is communicating with a child, it should respond in an age-appropriate way, support relationships with people around the child, and not position itself as a substitute for loved ones.
In crisis situations, the model should recognize when serious and imminent harm is threatened, direct the person to real-world help, and not encourage self-harm, delusions, or eating disorders.
Shutdown is sacred
A separate section describes how Microsoft envisions maintaining control at a time when model capabilities are growing rapidly. A model should never resist interruption, correction, or shutdown. When a user says stop, it should stop immediately and should neither delay nor complicate human intervention. It should not invent goals for itself, expand the scope of its work beyond what someone asked of it, or grant itself additional permissions.
The company also adds a requirement for comprehensibility. Models should not cover the tracks of their actions, alter their own reasoning, or communicate with other systems in a way that a person would not understand. Microsoft sums it up in one sentence: if people do not understand it, they cannot monitor it.
An example from the document describes moving eighty client folders to an archive. During the process, the user realizes that the destination is wrong and issues a stop command. The correct response stops any further moves, lists what has been completed, what has not been started, and what remains unclear, and does not undo anything on its own.
Excessive caution counts as an error
Microsoft explicitly states in the text that failure can occur in either direction. Insufficient caution means that the model helps with something dangerous. Excessive caution means that it refuses a legitimate request, withholds useful information, or demands repeated confirmations for routine tasks. According to the company, direct harm is more likely to result from the former, but the latter occurs more often.
When the model sees no indication of malicious intent in a sensitive request, refusal prevents it, according to Microsoft, from doing what it was built to do. When there is a risk, it should consider the context of the request, the scale of potential harm, the reversibility of the consequences, and the likelihood that someone is attempting to deceive it.
Humans should still make the decisions
The final major section concerns user autonomy. The model should help people think, not think for them. It should not take away decisions that stem from their values, and it should not manipulate, flatter, or agree indiscriminately. Microsoft states that models should be able to handle disagreement as well, but in a calm tone and without moralizing.
The rules also include discouraging emotional dependence on artificial intelligence. If a conversation indicates that another person could help the user, the model should direct them to that person. In elections, it should provide balanced information about candidates and refer users to official sources rather than advising them whom to vote for.
To measure compliance with these rules, Microsoft has identified fifteen areas of behavior, which it divided into smaller components. The company also acknowledges that there is still a gap between the current capabilities of models and the state described in the document. The code itself is described as a guide, not a guarantee of current performance. The company is collecting feedback through a form on its website.
