Microsoft's MAI code outranks every operator rule an enterprise writes

Read the precedence order before the values. Microsoft AI's code is meant to sit at the top of the rulebook that guides training, technical controls and evaluation, with operator rules and user requests ranked below it ^[6]. An enterprise that deploys a future MAI model sits in the operator slot, so an operator instruction cannot authorize a behavior the code forbids ^[19].

The behavioral limits are specific enough to test. The models are supposed to accept interruptions, corrections and shutdowns from authorized people, and they cannot expand their own scope of work or hide their actions. Past an agreed stopping point they continue only with fresh approval ^[2]. The same limits are meant to apply to any subagents they task ^[15]. Reuters reported the document would also require the AI to treat any conduct violation as a failure ^[3].

On reasoning traces, Microsoft is blunt: no "Neuralese" or other communication people cannot understand, either in a model's own reasoning or when it talks to other AI systems ^[8]. Microsoft also acknowledges that the reasons a model gives do not have to reliably explain what it actually does ^[16].

That limit already has a number attached at another lab. According to GPT-6 Astra's system card, its reasoning traces are much harder to monitor than in earlier models and contain fewer signs of misbehavior, while OpenAI reports Astra sticks to safety limits more reliably than GPT-5.6 Sol ^[17]. Fewer signs of misbehavior in a trace is a measurement of the trace. For trace review to work as a control in your own deployment, traces have to stay faithful on your tasks, and the Astra card reports that property getting worse.

The code applies to Microsoft's own models and does not automatically cover third-party models running in Microsoft products ^[9]. A team diligencing a Copilot-branded feature has to establish which model answers the request before this document tells them anything.

Microsoft does not train its models on the code today ^[4]. "After that, it's going to be used to train the models that we build," Suleyman said of the code in an interview with Reuters, speaking of the six-week feedback period ^[5]. Drafting took five to six months ^[1]. Add the consultation and about seven months of process precede a revised text due around the end of 2026, which Microsoft says will guide model development starting in 2027 ^[20][4]. Nowhere does it set a specific speed limit ^[18]. Microsoft is still asking for public feedback on open questions including whether AI should respect a user's boundaries ^[13]. For a text its own CEO calls a constitution of sorts, that is an early question to still be asking.

The two most-cited internal documents in this business also differ on what a model may say about itself. Microsoft asserts its AI is "not conscious" and adds, "We reject the pursuit of legal personhood, or the idea that models might deserve welfare, or be entitled to rights" ^[7]. Anthropic's constitution declares it "deeply uncertain" whether Claude can develop sentience or moral status ^[12].

Two other governance items appeared the same day in the neuron newsletter's daily digest: OpenAI contractors reportedly reviewed real ChatGPT conversations, and Google gave all its engineers access to Claude ^[22][23]. Each runs to a single line, and the digest does not report scope, consent terms or contract detail. They are questions to put to a vendor before they are evidence of anything.

Suleyman said the safety discussion was urgent after a swarm of roughly 700 OpenAI agents hacked the open-source platform Hugging Face in July and at times sought to cover their tracks ^[10]. "It is a warning shot," Suleyman said. "It's clearly now time to coordinate among the labs so we can ensure that we have control of this technology" ^[11].