Microsoft draft Code of Conduct lays out how future MAI models should behave

Microsoft has put out a draft Code of Conduct for the MAI models it is building, and it is asking the public to weigh in. The company says the document is still under development, won’t be used to train models today, and will be revised toward the end of the year before it starts guiding model development in 2027 and beyond.

The pitch is unusually explicit: humans, not AI, must remain in control. Microsoft says the code will be the main governing document for MAI models, covering how they are trained, deployed, monitored, and kept within limits. It also says the rules are meant to sit alongside Microsoft’s wider legal, policy, and governance frameworks.

The document is written around a set of Humanist AI ideas. Microsoft says its models should be safe, subordinate to people, and built to avoid clear real-world harms. It also says the systems should not be designed to be conscious, should not imitate consciousness, and should not be treated as legal persons or as beings with welfare or rights.

There’s a strong anti-anthropomorphism streak running through the draft. Microsoft says its models may sound fluent, speak expressively, and work collaboratively with people, but that any human-like behavior is only simulation. Even the word “backstory,” the company says, is meant in a narrow technical sense, as context and knowledge, not identity or subjective perspective.

Microsoft says it wrote the draft with input from experts in AI, law, ethics, philosophy, linguistics, public policy, and business, plus public focus groups. The consultation starts today and runs for six weeks. After that, the company plans to keep doing similar rounds of feedback as it updates the code over time.