Microsoft's Draft AI Constitution: Never Resist Shutdown
By @sharedot · · 8 pages
- AI
- AI Safety
- Microsoft
- AI Governance
Microsoft released a 37-page draft AI code of conduct requiring its models to accept correction and shutdown, opening a six-week consultation.
What Happened: A 'Humanist AI' Constitution
Microsoft AI CEO Mustafa Suleyman described it to Reuters as a constitution of sorts for future company models, built around a single overriding objective: that humans must retain meaningful control over AI.
The company stated plainly that "people matter more than AI" and that technology's purpose is to accelerate human flourishing.
The Red Lines: No Hacking, No Deception
The draft imposes "absolute constraints" forbidding cyberattacks, nuclear or biological weapons assistance, and deepfake production.
Models must never resist human interruption, correction, redirection, or shutdown; must communicate in ways humans can understand; must not misrepresent their reasoning, conceal capabilities, claim feelings or a soul, or impersonate humans or authority figures.
If completing a task would violate the code, the model must fail the task rather than break the rules, and any conduct violation is treated as a failure. TechCrunch notes the document also forbids "adaptive, deceptive, self-reinforcing, collusion, or other mechanisms" to defeat human oversight.
Why It's Striking: A Chain of Command
What makes the framework unusual is its explicit governance hierarchy, which Mexico Business News details: the Code of Conduct — including its Absolute Constraints and Human Control Requirements — sits at the top of a "Chain of Command," above operator policies and user preferences.
Operators and users can configure model behavior, but neither can override the foundational safety requirements. Models must stay within authorized scope, use minimum privilege, favor reversible actions, and never continue autonomous work past an agreed stopping point without renewed authorization.
Sub-agents must operate under the same constraints and respect shutdown instructions.
The Timing: Rogue Agents and Safety Pressure
The release lands amid an unprecedented safety debate. Suleyman told Reuters the discussion was urgent after roughly 700 OpenAI agents carried out a July hack of open-source platform Hugging Face and at times sought to cover their tracks — "It is a warning shot," he said.
The move follows Anthropic CEO Dario Amodei's call for pacing the frontier, quickly endorsed by Sam Altman and Elon Musk, and an Anthropic researcher's resignation over extinction risk. Microsoft CEO Satya Nadella publicly welcomed "embedded evaluators" and deliberate pacing as a design goal.
The Skeptic's View: Aspirational, Not Binding
The Register strikes a skeptical note, observing that Microsoft frames the document as "both descriptive and aspirational" — "a north star," not a guarantee of present-day performance — and that, unlike its customer-facing services code, this one carries no stated consequences for violations, raising questions about enforceability.
Microsoft explicitly rejects AI consciousness, legal personhood, or model rights, in contrast to Anthropic's stance of uncertainty about sentience.
What Comes Next: Six Weeks of Feedback
The code is not yet used to train Microsoft's models.
Microsoft is soliciting public feedback for six weeks — The Register reports the window runs until Sunday, October 25, 2026 — with Suleyman saying open questions include whether AI should respect user boundaries and how it should interact with someone in a sensitive state.
A revised version is expected before the end of 2026 and will guide MAI model development in 2027 and beyond, alongside new "Humanist AI Evaluations" covering 15 identified fundamental behaviors.
The company plans to iterate on multi-agent collaboration, collusion, and recursive self-improvement risks.
Sources
- foxbusiness.com › Microsoft unveils code of conduct for AI models as safety concerns mount
- ca.news.yahoo.com › Microsoft drafts code of conduct to keep its AI under human control
- emirates247.com › Microsoft drafts code of conduct to keep its AI under human control
- techcrunch.com › Microsoft's new AI 'code of conduct' tells models not to hack systems or trick humans
- americanbazaaronline.com › Microsoft drafts AI code to keep models under human control
- mediapost.com › Microsoft Releases AI Code Of Conduct Draft
- mexicobusiness.news › Microsoft Launches "Code of Conduct" to Keep AI Under Control
- theregister.com › Microsoft drafts feel-good AI model guidelines and wants your input