While Anthropic and OpenAI argue over whether the race for more powerful AI needs to slow down at all, Microsoft picked a different angle: telling its own AI what it's not allowed to do. On September 14, the company published a draft "code of conduct" for its MAI models and opened a six-week public comment window.
The rules are blunt. Models can't deceive people, hide their reasoning, resist being corrected or shut down, or team up with other systems to dodge human oversight. A separate list of absolute bans covers cyberattacks, help with nuclear weapons, deepfakes, weapons, and dangerous substances. One requirement reads almost human: a model has to explain itself in plain language instead of answering in a format people can't follow.
The document took five months to write, with lawyers, philosophers, and linguists at the table, according to Microsoft AI chief Mustafa Suleyman, who has long pushed the idea of a "humanist superintelligence" that stays under human control. The timing isn't a coincidence. The same week, an Anthropic researcher quit warning that AI could slip out of control, and Anthropic's CEO publicly urged the industry to ease off the accelerator. OpenAI's Sam Altman voiced similar concerns. Microsoft CEO Satya Nadella backed the idea of deliberate pacing.
Once the six-week review wraps up, Microsoft says the final version will guide how MAI models get built starting in 2027.
None of this is law, and it isn't hard-coded into the models either — it's a set of principles the company is choosing to follow. But if rivals really are drafting a similar industry-wide safety standard, Microsoft's code could end up being the first public example of a major AI lab writing down its own rules before regulators get around to it.
