Microsoft AI published a draft Code of Conduct for its MAI model family on 14 September and opened a six-week public consultation. The document lists constraints no model may break regardless of the prompt — from offensive cyber operations to mechanisms that would let it avoid being switched off. It is the first time Microsoft has exposed its own safety limits to outside criticism before enforcing them.
Key takeaways
- Draft published 14 September 2026, public consultation runs for six weeks.
- Five parts: Humanist AI, Safety, Operational Guidelines, Operational Defaults, Conclusion.
- Absolute constraints are split into frontier risks, public-safety risks and personal harms.
- Covers MAI models for transcription, image generation, coding, voice synthesis and reasoning.
- Satya Nadella publicly backed slowing the pace of development in favour of alignment work.
Four limits a model may never cross
Under frontier risks?Frontier risks: Hazards tied to the most capable models — those whose consequences reach beyond a single user and can affect entire systems or communities., Microsoft names four absolute constraints.
The four constraints the document marks as absolute — they hold regardless of what the prompt says:
A second group covers personal harms — deepfakes, child safety, human dignity and crisis response. The most interesting clause concerns self-preservation: MAI models may not use adaptive, deceptive or collusive mechanisms to evade oversight or shutdown.
People matter more than AI. AI should be a tool, not a person, and should never resist being switched off.
From the draft Code of Conduct for MAI Models, Microsoft AI.
Humanist superintelligence as the foundation
The document builds on humanist superintelligence, a concept Microsoft AI set out in November: systems that work solely for people, inside defined limits and under human control.
Three overarching objectives follow — keeping humans in control, drawing a clear line between model and person ("it is not conscious and should not be designed to imitate consciousness"), and human flourishing. A separate plural values principle requires support for differing worldviews within safety boundaries.
The third document of its kind
Microsoft joins Anthropic, with its constitution for Claude, and OpenAI, with its Model Spec. The difference is procedural: this is a draft put out for comment, because — as the company writes — "AI designed to serve humanity cannot be determined by only one company." Satya Nadella endorsed the direction, citing the need for "deliberate pacing" in alignment work, meaning a conscious slowdown in developing the most capable systems.
| Company | Document | Publication mode |
|---|---|---|
| Microsoft AI | Code of Conduct for MAI Models | Draft open for comment, 6 weeks |
| Anthropic | Constitution for Claude | Published document |
| OpenAI | Model Spec | Published document |
Why it matters
Value statements are cheap, enforceable constraints are not. This document's worth depends on whether Microsoft ties its clauses to concrete pre-release tests and whether anyone outside the company can verify them. What is genuinely new is the mode: instead of announcing a finished rulebook, Microsoft is exposing it to criticism before deployment. That shifts the weight from press release to process.
What's next?
- The consultation closes six weeks after 14 September, with feedback collected through a form on the Microsoft AI site.
- The text is marked as a draft, so the final version may differ from the published proposal.
