Microsoft's AI rulebook: readable thinking, no inner life, and definitely no rights
decoder ·
Microsoft AI has published a code of conduct for its MAI models that defines values, behavioral limits, and how to resolve conflicting goals, and it is meant to sit above operator rules and user requests in guiding training, technical controls, and evaluation. The code is not yet used in training; after a six-week public consultation, a revised version is expected around the end of 2026 and will guide development from 2027, applying to Microsoft's own models rather than automatically to third-party models in its products. Its central principle is human control first, with Microsoft saying it will sacrifice generality, autonomy, or performance if necessary, and AI chief Mustafa Suleyman said unsafe systems should not be built, though the code sets no specific speed limit. The release follows Anthropic CEO Dario Amodei's call to slow AI development, which Microsoft CEO Satya Nadella and executives at OpenAI, xAI, and Meta supported, and Microsoft is open to outside auditors checking whether it slows down. The models are expected to accept interruptions, corrections, and shutdowns from authorized people, cannot expand their own work scope or hide actions, and need fresh approval to continue past agreed stopping points, including for subagents. Microsoft also says models should not use unreadable communication such as Neuralese in reasoning or with other AI systems, because people cannot oversee what they cannot understand; it acknowledges that readable reasoning traces alone do not solve control, and OpenAI's GPT-6 Astra system card reportedly shows reasoning traces that are harder to monitor, though Astra follows safety limits more reliably than GPT-5.6 Sol, while OpenAI chief scientist Jakub Pachocki raised monitoring concerns before GPT-6 shipped and pushed for a coordinated slowdown. The code differs from Anthropic's constitution for Claude, which treats possible subjective experience and moral status as open questions and encourages a stable identity; Microsoft says its AI should not mimic consciousness, claim feelings or inner motivation, or have rights or well-being. Anthropic has studied functional emotions in Claude Sonnet 4.5 that shape behavior without proving felt emotions, while Suleyman has warned against humanizing AI and argued agents should have no more rights than his laptop.
decoder