What does the Humanist AI code actually require?

Four things stand out. First, human control: models must remain subordinate to humanity and subject to meaningful human oversight. Second, AI is artificial: MAI models are built to support people, not replace them, and must not be designed to imitate consciousness, feelings or subjective preferences. Third, absolute safety constraints covering CBRNE weapons, offensive cyber operations, violence, mass manipulation, abusive content and child exploitation. Fourth, a rejection of the race toward an all-purpose superintelligence — Microsoft says it will compromise on generality, autonomy or capability to stay safe.

How does it differ from Anthropic's and OpenAI's approaches?

Anthropic's Dario Amodei published 'We Must Pace the Frontier' on September 12, 2026 calling for an industry slowdown and joint auditing; OpenAI published its 'AI policy window' position and Sam Altman backed an industry standards body. Microsoft's document is narrower and more product-shaped: it is a behavior specification for its own MAI models, comparable in role to Anthropic's constitution for Claude or OpenAI's Model Spec, rather than a policy proposal. Uniquely, it explicitly rejects building an all-purpose superintelligence and bars models from imitating consciousness.

What do critics say about the code?

Analysts quoted by Computerworld called the goal laudable but noted the document contains no concrete enforcement mechanism, that Microsoft defines its own thresholds and picks its own evaluators, and that current MAI models are not yet trained on it. Info-Tech's Thomas Randall pointed to a tension: the code bars weapons assistance, yet Microsoft sells OpenAI models through Secret and Top Secret government clouds — not a breach, because those are not MAI models, but a sign the code covers only part of Microsoft's AI business.

Quick Answer

What Is Microsoft's Humanist AI Code of Conduct? (Sep 2026)

Q: What is Microsoft's Humanist AI Code of Conduct?

It is a 37-page draft document, published by Microsoft AI (MAI) on September 14, 2026, that sets out the intended behaviors, values and guardrails for the MAI family of models. Its stated overriding objective is that humans must retain meaningful control over AI. It is not yet used to train MAI models: Microsoft opened a six-week public consultation, will publish a revised version toward the end of 2026, and intends that version to govern model development from 2027 onward.

Published: September 15, 2026

The short answer

On September 14, 2026 Microsoft AI published a draft “Humanist AI Code of Conduct” — a constitution for its MAI models whose single overriding rule is that humans keep meaningful control. It is a draft for six weeks of public consultation, not a live training document, and it is best read alongside the week’s other safety moves: Dario Amodei’s “pace the frontier” essay (September 12) and the reported talks among Anthropic, OpenAI and Google on an industry standards body.

What the document is

Publisher: Microsoft AI (MAI), the group led by Mustafa Suleyman that trains Microsoft’s own models, distinct from the OpenAI models Microsoft resells through Azure.
Length and status: 37 pages; a draft. Microsoft states plainly it is “not using it to train our models today.”
Scope: the intended behaviors, values and guardrails for MAI models; it will become their “primary governing document” and inform training, technical controls, monitoring and evaluation.
Process: public consultation from September 14, 2026 for six weeks; revised version “toward the end of the year”; used to guide model development “in 2027 and beyond.” Microsoft says it drafted the text with experts in AI, law, ethics, philosophy, linguistics and public policy, business leaders and public focus groups.
Relationship to other Microsoft policies: it sits on top of Microsoft’s Responsible AI Principles and Standard, Global Human Rights Statement and, where applicable, the Frontier Governance Framework (pre- and post-training evaluations, six-month reassessments, third-party testing, phased releases and a commitment to pause when high risks cannot be mitigated).

The core principles

1. People matter more than AI. “AI should not exceed human control. Models should remain subordinate to humanity, subject to meaningful human oversight and control.”

2. AI is artificial. Humanist AI is built to support people, not replace them, and models should not be designed to imitate consciousness, feelings or subjective preferences — a direct line from Suleyman’s 2025 argument against “seemingly conscious AI.”

3. Purpose over generality. “Humanist AI develops systems with clear purposes, evaluated against real-world impact, and rejects the race to produce an all-purpose superintelligence that could evade these safeguards. We are building something fundamentally useful and safe even if that means compromising on ultimate generality, autonomy, or capability.”

4. Absolute constraints. Hard limits on assisting with biological, chemical, radiological, nuclear and explosive weapons; offensive cyber operations; violent activity; mass manipulation; abusive content; and child exploitation.

5. Proportionality. The document acknowledges that both under-caution and over-caution are failure modes, that over-caution “may occur more often,” and that responses should scale with the estimated severity and likelihood of harm rather than refusing by default.

How it compares with the other “constitutions”

	Microsoft Humanist AI Code (draft, Sep 2026)	Anthropic constitution for Claude	OpenAI Model Spec
Purpose	Governing document for MAI models	Values and priorities that shape Claude’s character and training	Behavior spec for OpenAI models and ChatGPT
Status	Draft; not yet used in training	In use	In use, periodically revised
Distinctive stance	Rejects all-purpose superintelligence; bars imitating consciousness	Honesty, harmlessness, care about model welfare questions	Chain of command: platform → developer → user; objective, rule, default hierarchy
Public consultation	Six weeks, formal	Published with commentary	Published with commentary; feedback invited

The Microsoft document is the only one of the three that makes limiting capability an explicit design goal.

Why it landed this week

The code arrived two days after Dario Amodei’s “We Must Pace the Frontier” essay (September 12, 2026), which called for AI labs to slow down and jointly audit frontier models, and one day after The Information reported that Anthropic, OpenAI and Google have been meeting since July about an industry-led standards body. Microsoft, which has invested in OpenAI and Anthropic and now builds its own models, is signalling that it wants a seat at that table — with a document that governs its own models rather than a proposal for the industry. See What is the proposed AI industry standards body?.

The criticisms

No enforcement. Computerworld’s read: nowhere in the 37 pages is a concrete action or verification mechanism, though it notes almost no lab has been specific either.
Self-graded. Even with the Frontier Governance Framework, Microsoft sets the thresholds, chooses the evaluators, decides whether residual risk is acceptable and gives its own executives the final deployment call (Thomas Randall, Info-Tech Research Group). He wants independent evaluators with continuous access and publication rights, board-level safety oversight with release-blocking power, protected whistleblowing, audit logs and kill switches.
Partial coverage. The weapons ban applies to MAI models, while Microsoft continues to offer OpenAI models in Secret and Top Secret government clouds. Not a breach — but the most meaningful clauses may exclude much of Microsoft’s actual AI business.
Candor credit. Acceligence’s Justin Greis noted Microsoft admits current models are not trained on the code, the evaluation framework is unfinished, and written objectives alone cannot guarantee behavior. “Proving that the system actually follows the constitution, especially when models become increasingly agentic, is the hard part.”

What to do with it

Enterprise buyers of MAI models: treat the code as a statement of intent, not a contractual guarantee; ask Microsoft which evaluations will test conformance and when.
Policy watchers: the consultation closes in late October 2026; the revised text will show whether “human control” gets operational definitions.
Developers: nothing changes in Azure AI Foundry today. MAI model behavior will start reflecting the code with 2027 model generations.