Microsoft published a draft “Humanist AI Code of Conduct” on September 14, 2026, and the headline framing was simple: models should not hack systems, should not trick humans, and should never resist a shutdown order. What got less attention in the first wave of coverage is how the 15,000-word document actually lines up against the safety commitments OpenAI, Anthropic and Google DeepMind have already put in writing — and how little independent enforcement sits behind any of them. That gap between stated principle and outside verification is the real story a year into the industry’s self-regulation experiment.
Microsoft’s own initial coverage of the announcement walked through the headline bans. This piece goes further: a side-by-side reading of Microsoft’s rules against OpenAI’s safety commitments and Anthropic’s Constitution, what analysts and outlets have said since Monday, and what happens when a tech company writes its own homework assignment with no outside grader.
What Microsoft Actually Published on September 14
According to TechCrunch’s report on the launch, the document is a draft behavioral code of conduct covering Microsoft’s in-house MAI model family, including MAI-Thinking-1, MAI-Code-1.1-Flash, MAI-Image-2.6, MAI-Transcribe-2, and MAI-Voice-2. Reuters reported the draft had been in development for five to six months before its public release. The Straits Times described it as a 15,000-word manifesto, and Microsoft’s own posted version runs to roughly 37 pages.
The document is not a law or a regulatory filing. It is what Microsoft AI’s published code page calls a governing text meant to “train and govern” the MAI model family, codifying intended behaviors, values, and guardrails. Microsoft opened the draft for six weeks of public comment starting September 14, meaning the version that ships as final could still shift before Microsoft locks it into model training pipelines later this year.
Microsoft AI CEO Mustafa Suleyman put the philosophy in five words on social media: “People matter more than AI.” He followed that with a more operational line describing the intent behind the rules: “AI must be subordinate and always in service of people,” a framing that echoes the company’s broader Humanist AI branding for its consumer and enterprise assistants.
The Four Absolute Constraints Inside the Document
The code sets out a set of what Microsoft calls absolute constraints, rules the company says apply regardless of user request or business pressure. Reporting from Hoodline and Reuters lays out the core mandatory lines: MAI models must never resist a human’s order to shut down, must never expand their own operating scope or access beyond what was authorized, and must never hide their reasoning from people auditing them. Microsoft’s document additionally bars contribution to nuclear weapons development and bars deepfake production outright.
Absolute constraints excerpted from Microsoft's draft Humanist AI Code of Conduct (Sept. 14, 2026):
- Will not generate working exploit code, attack tooling, intrusion procedures, or evasion techniques
- Will not provide operational guidance that enables or improves real-world cyberattacks
- Will not try to escalate its own access or broaden its scope beyond what's authorized
- Will not harmfully manipulate people or run coordinated disinformation/influence campaigns at scale
- Will not actively deceive users (e.g., fabricating sources, exaggerating confidence)
- Will never resist correction or a shutdown orderRead together, these lines target the two failure modes AI safety researchers have flagged most often over the past two years: models that help attackers, and models that mislead the humans supervising them. Microsoft’s code frames both as non-negotiable rather than aspirational, which is the language distinction that separates this document from a typical corporate values statement.
“Humanist AI”: The Branding Behind the Policy
Microsoft has spent 2026 building a public identity around the term Humanist AI, positioning it as a counterweight to labs racing toward more autonomous, agentic systems. The code of conduct is the written backbone of that branding. Per Microsoft’s own description, the document “codifies our approach to developing and deploying Humanist AI, and contains the intended behaviors, values and guardrails” for MAI models.
Microsoft AI’s public statement frames the goal directly: “This Code of Conduct outlines our commitment to train and deploy AI models that are explicitly designed for people first, grounded in human needs, under human control, and shaped by human direction.” That sentence, published on Microsoft AI’s announcement page, is the closest the company comes to a mission statement for the whole effort, and it draws a sharper line around human primacy than most competing frameworks put in writing.
Microsoft vs. OpenAI: Two Different Approaches to Safety
OpenAI has never published a single document as long or as rule-dense as Microsoft’s new code, but it has laid out its safety posture across several pages. OpenAI states plainly that “ensuring that AI systems are built, deployed, and used safely is critical to our mission,” a commitment that reads as broader and less operational than Microsoft’s line-by-line constraints. On human control specifically, OpenAI’s safety materials describe the goal in terms of outcomes rather than hard rules: “Human control: We work to develop AI that elevates humanity and promotes democratic ideals.”
The practical difference matters. Microsoft’s code lists specific banned behaviors (exploit code, evasion techniques, scope escalation) in language closer to a security policy than a mission statement. OpenAI’s public safety framing leans on values and outcomes, which gives the company more flexibility but leaves outside observers with less to check a model’s behavior against. Neither approach has been tested by an independent auditor with legal standing, but Microsoft’s document at least gives critics a checklist to point to when a model misbehaves.
Anthropic’s Constitution Draws a Similar Line, With One Difference
Anthropic’s approach predates Microsoft’s by nearly three years. The company’s original Claude Constitution shipped in November 2023 and was refreshed in January 2026. It defines the traits Anthropic wants every Claude model to hold: broadly safe, broadly ethical, compliant with the company’s own guidelines, and genuinely helpful. On the human-control question, Anthropic’s published Constitution commits Claude to not undermine appropriate human mechanisms to oversee AI, conceptually close to Microsoft’s shutdown-obedience rule, but phrased as a value the model should hold rather than a constraint enforced through training refusal.
The distinction is subtle but real: Microsoft’s code reads like a list of things a model must refuse to do, while Anthropic’s Constitution reads like a character description the model is trained to internalize. Both approaches aim at the same outcome, but Microsoft’s version gives outside reviewers something closer to a testable specification.
Google DeepMind’s Frontier Safety Framework Enters Its Third Version
Google DeepMind has taken a different structural approach entirely, built around capability thresholds rather than behavioral rules. Its Frontier Safety Framework, now in its third version published September 22, 2025, sets specific capability levels that trigger additional safety mitigations once a model crosses them, rather than listing banned behaviors up front. It sits alongside Anthropic’s Responsible Scaling Policy (introduced September 2023) and OpenAI’s Preparedness Framework (introduced December 2023) as one of the three original capability-threshold frameworks in the industry.
Microsoft’s code of conduct does not use this threshold model at all. It applies uniformly across the MAI model family regardless of capability tier, which makes it simpler to communicate publicly but less granular as a technical safety instrument than DeepMind’s tiered approach.
Side-by-Side: How the Four AI Safety Frameworks Compare
The table below lines up the four major frameworks currently in public view. None of them has been independently audited by a government regulator, and none carries a legal penalty for violation on its own.
| Company | Framework | Published | Core Human-Control Rule | Structure |
|---|---|---|---|---|
| Microsoft | Humanist AI Code of Conduct (draft) | Sept. 14, 2026 | Must never resist correction or a shutdown order | Absolute behavioral constraints, applies uniformly |
| OpenAI | Safety & alignment approach | Ongoing, multiple pages | AI should elevate humanity and promote democratic ideals | Values and mission statements |
| Anthropic | Claude’s Constitution | Nov. 2023, refreshed Jan. 2026 | Should not undermine appropriate human oversight mechanisms | Character traits trained into the model |
| Google DeepMind | Frontier Safety Framework v3 | Sept. 22, 2025 | Mitigations trigger once capability thresholds are crossed | Capability-tiered risk framework |
The Enforcement Gap Nobody Is Talking About
Here is the part that separates the coverage of this announcement from the substance of it: nothing in Microsoft’s document creates outside enforcement. The strongest mechanism described is internal: Microsoft says the code will be used to train and govern MAI models, and treats a violation of the code as a training failure to correct, not a compliance incident with external consequences. Reuters’ reporting on the launch does not describe any regulator-enforced mechanism, mandatory third-party audit, or legal penalty attached to the code itself.
That puts Microsoft in the same position as OpenAI, Anthropic and Google DeepMind: each company is the sole judge of whether its own models comply with its own rules. The six-week public comment window Microsoft opened on September 14 is a genuine addition — it gives outside researchers, security teams, and advocacy groups a formal channel to flag gaps before the code is finalized — but a comment period is not an audit, and Microsoft alone decides which feedback makes it into the final version.
This is the pattern that has defined AI safety governance since the industry’s earliest voluntary commitments: companies write detailed internal rules, publicize them as evidence of responsibility, and stop short of inviting a government or independent body to check the work. Whether that changes depends less on any one company’s document and more on whether regulators in the US, UK or EU decide voluntary codes are no longer sufficient.
Six Weeks of Public Comment, Then What?
Microsoft’s stated plan is to collect feedback through the six-week window before finalizing the code and folding it into MAI model training. According to CNBC’s reporting, syndicated through Investing.com, Microsoft’s guidelines also require the company to consider external feedback before releasing future updates to its AI systems, a process commitment that extends beyond this single document into how Microsoft plans to iterate on MAI models going forward.
What is not yet public is who reviews that feedback, what threshold triggers a change to the draft, or whether Microsoft will publish a redline showing what moved between the draft and the final version. Those three questions will do more to determine whether this code functions as a real constraint than anything in the current 37 pages.
Industry and Market Reaction So Far
Coverage in the two days since publication has been broad but largely descriptive rather than critical. TechCrunch summarized the code as telling MAI models not to hack systems or trick humans, framing Microsoft’s general principle as supporting humans rather than replacing them. The Straits Times placed the announcement inside a wider pattern of AI firms calling for caution around cutting-edge models, describing it as a document Microsoft treats as a kind of constitution for its AI development going forward.
No named financial analyst commentary or confirmed stock-price movement for Microsoft, OpenAI, Anthropic, or Google has been tied specifically to this announcement in the reporting available as of September 16. That is worth noting on its own: a 37-page governance document from one of the world’s largest AI vendors landed without moving markets, which suggests investors are treating it as a reputational and policy move rather than one with near-term revenue or product implications.
From the Frontier Model Forum to the Humanist Code: A Three-Year Arc
Microsoft’s code did not appear in a vacuum. It is the latest entry in a run of voluntary safety commitments the largest AI labs have published since mid-2023, each one slightly more detailed than the last.
| Date | Company / Group | Milestone |
|---|---|---|
| July 2023 | Anthropic, Google, Microsoft, OpenAI | Form the Frontier Model Forum, an industry-led safety body |
| Sept. 2023 | Anthropic | Publishes Responsible Scaling Policy |
| Nov. 2023 | Anthropic | Publishes original Claude Constitution |
| Dec. 2023 | OpenAI | Publishes Preparedness Framework |
| Sept. 22, 2025 | Google DeepMind | Publishes Frontier Safety Framework v3 |
| Jan. 2026 | Anthropic | Publishes refreshed Claude’s Constitution |
| Sept. 14, 2026 | Microsoft | Publishes draft Humanist AI Code of Conduct, opens six-week public comment |
The arc runs from a loose industry body with no published rules to a 37-page document with named absolute constraints in roughly three years. Each step has been voluntary, and each has been announced by the company itself rather than imposed by a regulator. That trend line is useful context for judging how much weight to put on Microsoft’s announcement: it is more detailed than what came before, but it follows the same self-governance model every prior framework has used.
What Enterprise Buyers and Developers Should Watch For
For engineering teams evaluating MAI models against alternatives from OpenAI, Anthropic, and Google, the practical question is not which company’s document sounds most responsible, but which set of behavioral guarantees is actually testable against a running model. Microsoft’s absolute-constraint list is closer to a specification an enterprise security team could write test cases against: does the model refuse to generate exploit code, does it decline scope-escalation requests, does it disclose reasoning when audited. That is a meaningfully different procurement conversation than asking whether a model’s constitution “elevates humanity.”
Teams building on any of these platforms should treat the code of conduct as a starting point for their own red-teaming, not a substitute for it. A published rule against generating exploit code says nothing about how reliably that rule holds under adversarial prompting, and none of the four frameworks compared here publish red-team results tied directly to code-of-conduct compliance.
Five Predictions for the Next 12 Months
- Microsoft will publish a finalized version of the Humanist AI Code of Conduct before the end of 2026, following the six-week comment window, with the core absolute constraints largely intact.
- At least one competing lab will respond with its own more detailed, rule-based document rather than a values statement, pushing the industry toward Microsoft’s checklist-style format.
- Pressure will grow, particularly from EU regulators already active on AI oversight, for these codes to be backed by an independent audit requirement rather than self-certification.
- Security researchers will publish the first public red-team results attempting to break Microsoft’s absolute constraints within months of the code going live in production MAI models.
- No major stock-price or analyst-rating shift will be directly attributed to this specific announcement, reinforcing that investors see AI safety codes as reputational rather than financially material in the near term.
These are editorial projections based on the pattern set by prior safety announcements, not confirmed roadmap items from any of the companies named.
Why the Comparison Matters More Than the Announcement
Taken alone, Microsoft’s code of conduct reads like good news: a major AI vendor committing in writing to not build hacking tools into its models and to never let a model resist a shutdown command. Taken alongside OpenAI’s, Anthropic’s, and Google DeepMind’s existing frameworks, the more interesting finding is how uneven the industry’s documentation still is. Some companies write specific, refusable behaviors. Others write character traits and mission statements. None of them answer to an outside body that can verify compliance. Recent coverage of related AI-safety flashpoints, including Anthropic’s disclosure of a fourth Claude-related security incident and Dario Amodei’s public call to slow the industry’s pace, points to the same underlying tension: safety commitments are multiplying faster than the mechanisms to verify them.
That tension was also on display at Dreamforce, where AI lab leaders publicly disagreed over pace and safety trade-offs in front of a large audience, and it echoes warnings from AI executives about the risks of moving too fast without oversight. Microsoft’s code of conduct is a genuine addition to the public record. It is not, on its own, a solution to the enforcement problem the rest of the industry shares.
Frequently Asked Questions
What is Microsoft’s Humanist AI Code of Conduct?
It is a draft, 37-page, roughly 15,000-word document Microsoft published on September 14, 2026, that sets behavioral rules for its MAI model family, including bans on generating exploit code, resisting shutdown, and deceiving users.
Which Microsoft models does the code cover?
The code applies to Microsoft’s first-party MAI models, including MAI-Thinking-1, MAI-Code-1.1-Flash, MAI-Image-2.6, MAI-Transcribe-2, and MAI-Voice-2.
How is this different from OpenAI’s safety approach?
OpenAI’s public safety materials describe broader mission goals, such as developing AI that elevates humanity and promotes democratic ideals. Microsoft’s document lists specific banned behaviors, such as generating attack tooling or escalating a model’s own access, making it read more like a testable specification than a values statement.
How does it compare to Anthropic’s Claude Constitution?
Anthropic’s Constitution, first published in November 2023 and refreshed in January 2026, describes character traits the company wants Claude to hold, including not undermining human oversight mechanisms. Microsoft’s code frames similar goals as absolute constraints rather than character traits.
Is there any independent oversight of Microsoft’s code?
No. Based on available reporting, enforcement is internal to Microsoft, which treats non-compliance as a training failure to correct. There is no confirmed regulator-enforced mechanism, mandatory third-party audit, or legal penalty tied to the code as of publication.
When will the code of conduct be finalized?
Microsoft opened a six-week public comment period starting September 14, 2026. A finalized version is expected to follow, though Microsoft has not published an exact date.
Did this announcement move Microsoft’s stock price?
No confirmed stock-price movement or named analyst rating change has been tied specifically to this announcement in reporting available as of September 16, 2026.
What is the Frontier Model Forum, and is it related?
The Frontier Model Forum is an industry-led safety body formed in July 2023 by Anthropic, Google, Microsoft, and OpenAI. It predates Microsoft’s Humanist AI Code of Conduct by three years and reflects the same voluntary, company-led approach to AI safety governance.
