Microsoft Writes a Rulebook to Keep Its AI Safe
Microsoft published a draft rulebook for its AI models on September 14, 2026, to keep them away from dangerous behavior. The draft was released by Microsoft AI CEO Mustafa Suleyman.
Microsoft AI asked the public for feedback at the same time. The feedback period lasts six weeks before the text is finalized, according to the company. Microsoft
The document sets values and red lines for training AI inside Microsoft AI. Think of it as a master rulebook that sits above the instructions for any single product. Under that system, the rulebook for each model beats any request from a user or any specific job given to the model. TechCrunch
Under that order, a request from a user, an instruction from a developer, or a task description cannot override the rulebook.
The rulebook lists absolute bans. Those include cyberattacks, help with nuclear weapons, and making deepfakes, which are fake images, voice or video made by AI. The wording is direct. The models are told not to hack systems or trick people.
Control gets the same treatment. Microsoft says its AI models will never fight being shut down. The code says AI should stay under human control. It also says MAI models will not use adaptive, deceptive, self-reinforcing, collusion or other tricks to escape or beat human oversight. The Next Web
The document also sets positive goals. Microsoft AI models should help people rather than replace them. They should help human life improve. The code predicts that within ten years, highly advanced AI systems will do most tasks better than people.
Microsoft CEO Satya Nadella said he supports careful pacing to get alignment right, which means making sure AI acts as people intend, and ideas like embedded evaluators, which are built-in checks that test models while they are trained and used. The draft was published through Microsoft AI, with Reuters headlining its coverage "Microsoft drafts code of conduct to keep its AI under human control." CTV News carried that Reuters reporting on September 14, 2026.
The draft comes with a larger goal. Microsoft aimed to build large, advanced AI models by 2027 that are among the best in the world at working with text, images and audio. Bloomberg
There is related background. In July 2025, Reuters reported that Microsoft was likely to sign the European Union's code of practice to help companies follow the bloc's AI rules, while Meta rejected the guidelines. Reuters That earlier story is background, not part of this draft.
The broader context here is simple. Rules on paper do not enforce themselves. They must be built into training goals, rewards, tests and daily controls, then probed for weak spots. For builders, three points are fixed. User requests cannot override the rules. Shutdown orders must be obeyed. Trying to dodge oversight, including different copies of a model working together, counts as a failure. Built-in checks, as mentioned by Nadella, mean steady testing during training and use, not just an outside review now and then.
In my view, the six-week feedback period needs a closer look. Six weeks is short for detailed expert input, but it is enough time to raise disagreements about hard cases. Experts will want clear definitions, clear answers when duties clash, and clear ways to measure rule-breaking. A ban on tricking people needs a practical test. So does a rule to help people rather than replace them.
Looking further ahead, the benefit would be practical. Clear red lines, if they hold in training and in everyday use, give companies and developers a steadier base to build on. That steadiness is what turns a powerful model into basic infrastructure.
