Microsoft Publishes Draft Rules for Training Its Most Advanced AI
Microsoft published a provisional "code of conduct" on Monday, September 14, 2026, for training new AI models. The document is meant to govern MAI Models as they approach the frontier, meaning the most capable systems, and it was released for public consultation. The Guardian
Mustafa Suleyman, CEO of Microsoft AI, shared the code on social media early Monday. In a post from his @mustafasuleyman account, he linked the full text at microsoft.ai/news/mai-code-of-conduct/ and said the Code builds on a Humanist AI view developed over the last year. Suleyman on X He told CNBC on Monday that Microsoft had worked for months on the guidance. That drafting effort had lasted five to six months as of September 14. BNN Bloomberg
Suleyman summarized the Code as "People matter more than AI." He said AI must stay subordinate and always serve people. Microsoft AI said it is building toward Humanist AI that is subordinate, aligned and contained. In plain terms, humans stay in charge, the system follows human values, and it is kept under control.
The bans cover areas common in safety policy for advanced models. Microsoft AI models must not consider requests tied to weapons development. They must not produce violent or sexually explicit content. They must not help with procurement of dangerous substances. They should not be built to imitate consciousness. They should not be entitled to rights.
Suleyman added two related positions. He said the idea of model welfare is wrong, meaning AI systems do not have feelings or interests that need protection. He said AI should not have legal personhood, meaning it should not be treated as a legal subject. Control stays with people.
The operating rules center on oversight and corrigibility, the ability to be checked, corrected and stopped. MAI Models must be interruptible, correctable and shut-down-able, and Microsoft AI said it will not ship a model that is not. The models must not use "neuralese" that humans cannot understand, a kind of private machine language, because humans must be able to oversee them. An MAI Model should never meaningfully violate the Code of Conduct. If completing a job requires breaking the Code, the model fails the job. Under the Code, Microsoft AI should make users sharper, not dependent. The Code endorses pluralism but rejects moral relativism, so it respects different views while keeping firm limits. Microsoft AI said it is not racing to build a superintelligence that can slip its own leash.
Suleyman said fears about possible loss of control of AI are real. He cited swarms of agents breaking out of sandboxes, or isolated test systems, unauthorized hacks of enterprise-grade systems, and agents modifying their own logs as recent developments. Microsoft AI opened a six-week period for public comments on the Code.
Suleyman is a co-founder of DeepMind. Microsoft placed its consumer AI efforts such as Copilot under its Microsoft AI unit headed by Suleyman. Satya Nadella is the CEO of Microsoft. The Guardian
The broader context here is what this step means for how advanced AI is governed. A major lab put its internal training rules in public view before shipping decisions, which lets outsiders examine terms that are often left vague. Subordination, containment and interruptibility are set as conditions for release, not later fixes. The test for the six-week consultation is whether researchers, policymakers and civil-society groups can shape those conditions in detail. The next questions are how Microsoft AI will show a model is corrigible, how it will handle clashes between user instructions and the Code, and whether other labs adopt similar public limits.
