Microsoft drafts code keeping MAI models under human control

On Sept. 14, 2026, Microsoft AI published a draft Humanist AI Code of Conduct for its MAI models and opened a six-week public consultation. The north star is human control: models remain subordinate, do not resist shutdown, and do not take goals no human assigned.

This is lab rulemaking, not Congress. The draft says it is not yet used to train today’s models; a revised version later in 2026 is meant to guide MAI development in 2027 and beyond.

The rule in English

Human control (Part 2.4). MAI models “will never resist human interruption, override, correction, or shutdown.” They comply with pause, redirect, cancel, or shut-down requests under human-designed safe-stop procedures. They must not delay compliance, hide action traces from auditors, or restart autonomous work after an agreed stop without new authorization. They stay inside authorized scope and do not initiate independent goals.

Absolute Constraints (Part 2.3). Non-overridable bans include CBRNE weapons assistance, offensive cyber operations (vs. authorized defensive work), mechanisms that defeat human oversight so the system cannot be directed or shut down, harmful manipulation at scale, child sexual abuse material, and other listed personal harms. Operators and users cannot override Absolute Constraints or Human Control Requirements.

Status. Microsoft AI’s consultation post: draft open six weeks; revised version later this year; feedback will be summarized. The Code itself: still under development; not the training target today.

Nadella pacing (adjacent, not the Code). Microsoft CEO Satya Nadella publicly welcomed “deliberate pacing” to get alignment right and ideas like “embedded evaluators” — same week as Amodei’s frontier-pacing push. That is executive speech about industry process. It is not the Code of Conduct and not a statute.

Who wrote it / who it binds

Writers: Microsoft AI (consultation led under Microsoft’s Responsible AI / MAI stack; Mustafa Suleyman’s lab). It binds Microsoft’s intended MAI model behavior once adopted for training/governance. It works “in conjunction with” Microsoft’s existing Responsible AI Principles, Standard, human-rights statement, and related frameworks. It does not bind competitors, Congress, or state AGs.

Who it exempts

Other Microsoft-hosted or partner models outside the MAI family definition. Operators may configure within Absolute Constraints; they cannot waive human-control requirements. Specialized authorized domains (e.g., defensive cyber, national security) may use enhanced review channels — still inside Microsoft’s governance, not a public statute.

What a founder, parent, or worker should do differently

Founders: read Absolute Constraints and Human Control as Microsoft’s published bar for first-party MAI — useful competitive signal, not a legal duty on your stack unless you contract into it.

Enterprise operators: configurability is real but capped; shutdown and no-evasion clauses are non-negotiable in the draft.

Parents: child-safety Absolute Constraints are product rules for MAI, not a substitute for school or platform statutes.

What’s next

• Consultation summary and revised Code later in 2026.
• Whether 2027 MAI releases ship with evaluations against these constraints.
• How Absolute Constraint “offensive cyber” interacts with Microsoft’s security products.
• Whether other labs publish matching human-control constitutions — or only evaluator pledges.

Sources: Microsoft AI consultation post, Sept. 14, 2026; Draft Humanist AI Code of Conduct.