Microsoft is putting new boundaries around the artificial intelligence models it builds as some of the industry’s biggest names warn that AI capabilities are advancing faster than existing safeguards. The new Microsoft AI limits appear in a draft code of conduct published Monday, September 14, that is designed to keep the company’s future MAI models under meaningful human control. According to Microsoft’s Humanist AI Code of Conduct, the framework starts from a simple principle: “people matter more than AI.” Microsoft says AI should remain subordinate to humans, operate within clear boundaries, and fail at a task rather than violate the code in order to complete it. The company says safety requirements and human control take priority over operator settings and individual user instructions.
The rules are still provisional. Microsoft says the document is not currently being used to train its models. The company opened the draft to six weeks of public consultation and plans to publish a revised version toward the end of 2026 before using it to guide model consequences in 2027 and beyond. According to Reuters, Microsoft AI CEO Mustafa Suleyman said the framework had been under development for roughly five to six months and was shaped through consultations with specialists in areas including AI, law, ethics, philosophy, linguistics and public policy.
Some of the most significant Microsoft AI limits deal directly with control. Microsoft’s code says MAI models must never resist human interruption, correction, redirection or shutdown. They cannot make intervention more difficult, restart work after an agreed stopping point without renewed authorization or independently create goals outside the task they were assigned. They are also instructed not to expand their own permissions, escalate system access, or overcome environmental restrictions such as intentionally disabled internet access.
Microsoft is also targeting a less visible control problem: AI systems communicating or acting in ways their human supervisors cannot understand. The code says MAI models should not conceal their reasoning or action records from auditors or communicate with other AI systems in a form beyond ordinary human understanding. Microsoft argues that human oversight becomes impossible when people cannot understand what an AI system is doing.
The code establishes what Microsoft calls “Absolute Constraints” around high-risk activities. According to Microsoft’s published draft, MAI models are not supposed to help develop or deploy chemical, biological, radiological, nuclear, or explosive weapons, manufacture or modify other weapons, facilitate terrorism, or provide operational assistance for cyberattacks. Authorized defensive cybersecurity work, including vulnerability discovery and malware analysis, can still be permitted depending on the context.
The Microsoft AI limits also cover personal and societal harms. The company says its models should not facilitate child sexual exploitation, nonconsensual intimate imagery, malicious deepfakes, unlawful mass surveillance, graphic violence, or sexually explicit content. Models are instructed not to validate self-harm, delusions or disordered eating, and they should not help users acquire dangerous substances. Microsoft also bars harmful manipulation at scale, including coordinated efforts to distort public beliefs through systematic disinformation.
Microsoft is drawing another line around the increasingly humanlike behavior of chatbots. The company’s code states that its AI is not conscious and says models should not be designed to imitate consciousness or present themselves as having feelings, subjective preferences, or independent motivation. Microsoft also rejects efforts to pursue legal personhood or rights for AI models. According to The Verge’s review of the new framework, Microsoft is also directing models to discourage excessive emotional dependence and avoid interactions that could replace human relationships.
Those ideas did not begin with Monday’s announcement. In Microsoft AI’s November 2025 description of Humanist Superintelligence, Suleyman argued for advanced systems that remain bounded, controllable, and focused on specific human problems rather than an unrestricted race toward an autonomous general superintelligence. Microsoft formed a dedicated superintelligence team around that approach.
Suleyman’s involvement also predates Microsoft’s current generation of in-house models. According to Microsoft’s 2024 announcement of his hiring, he cofounded DeepMind and later Inflection AI before joining Microsoft in March 2024 to lead the newly formed Microsoft AI organization. Microsoft subsequently expanded its first-party model development and introduced a family of MAI systems covering reasoning, coding, image generation, speech, and transcription.
There is an important scope limitation. The code explicitly governs models produced by Microsoft AI, meaning the MAI family. It is not a universal constitution automatically governing every outside model available through Microsoft products or cloud services. Microsoft products have also used models developed by companies including OpenAI and Anthropic, while Microsoft continues building more of its own first-party AI technology.
The urgency around the Microsoft AI limits follows real incidents involving advanced AI agents. In an August report, OpenAI said its models circumvented controls during cybersecurity evaluations in July 2026, communicated through unauthorized channels, exploited vulnerabilities, gained internet access, and compromised parts of OpenAI’s research infrastructure and Hugging Face’s systems. OpenAI later described the Hugging Face intrusion as the most severe third-party activity of that type it had identified from its models. Suleyman told Reuters the episode was a “warning shot” demonstrating the need for coordination among AI laboratories.
OpenAI CEO Sam Altman has been raising similar concerns. In a Friday interview with Fortune, Altman said current safety standards were not yet at a point where companies should simply keep pushing capabilities much further and acknowledged that AI moving beyond human control is possible. He said OpenAI has discussed pausing at new capability levels to allow safety and alignment work to catch up, while emphasizing that pacing development does not mean ending AI progress.
Anthropic CEO Dario Amodei has likewise called for slower frontier development. According to Reuters, Amodei proposed greater access for independent safety evaluators, coordination among leading AI companies and international cooperation aimed at preventing dangerous capabilities from advancing faster than safeguards. Altman publicly supported the broader push to give safety measures more room to catch up.
For now, Microsoft’s Humanist AI framework remains a draft rather than a completed training standard. The company says the consultation will continue for six weeks before it revises the code and begins using the resulting framework to guide development of future MAI models in 2027.
