Microsoft AI publishes its constitution: "people matter more than AI"
Microsoft AI is putting up for public consultation a Code of Conduct that defines what its MAI models must do — and above all, must never do.
In brief
Microsoft AI has published a draft "Humanist AI Code of Conduct," a governance document meant to become the primary charter for its MAI models. It stakes out a clear thesis: AI must remain a subordinate tool, non-conscious, without rights or legal personhood, and incapable of resisting a human shutdown. The text does not yet train any model: it's open for comments for six weeks before a revision that will guide development in 2027.
🍺 Bar-stool version
Microsoft AI just published a sort of house rulebook for its future models, and the main message fits in one line: AI is a tool, it has no consciousness, no rights, and above all it has no right to argue when someone hits the stop button. The text also bans the model from hiding its reasoning or talking to itself in gibberish nobody understands — which, let's be honest, is a clause we'd love to slip into a few employment contracts. Small detail: this document isn't training any model today, it's up for consultation for six weeks and will serve as a compass "in 2027 and beyond," which is the elegant way of saying we'll see. It still matters, though, because declaring that a machine has no inner life is both a deliberate philosophical stance and a very practical way to dodge awkward questions the day it starts saying otherwise.
Key takeaways
- 1
The document is explicitly a draft: it's not being used to train models today, and a revised version is expected by year's end to guide development "in 2027 and beyond."
- 2
The public consultation runs six weeks and draws on experts in AI, law, ethics, philosophy, linguistics and public policy, plus general-public focus groups.
- 3
Microsoft AI flatly rejects legal personhood for models, the idea of AI welfare or rights owed to them, and bans any imitation of consciousness.
- 4
A "Chain of Command" ranks Code of Conduct > Operator policies > User preferences, with Absolute Constraints that no one can override.
- 5
The Human Control Requirements forbid the model from delaying a shutdown, hiding its reasoning traces, or communicating in incomprehensible "neuralese."
- 6
The text treats over-caution as a real failure mode, on par with under-caution, and calls for a response proportionate to the severity and reversibility of the risk.
- 7
15 core behaviors were identified to build evaluations, illustrated by 9 scenarios comparing aligned vs. misaligned responses, generated with MAI-Thinking-1.
A deliberate draft, not a product
Microsoft AI takes care to defuse any marketing-style reading: this Code of Conduct "is still under development" and is therefore "not used to train our models today." It's published for consultation, for six weeks, ahead of a revised version by year's end.
It's that revised version that will guide model development "in 2027 and beyond." The document describes itself as "both descriptive and aspirational" and acknowledges "a gap between current default behaviors and the full future scope of Humanist AI."
The text stresses it doesn't replace anything: it works alongside the Responsible AI Principles, the Responsible AI Standard, the Global Human Rights Statement and, where applicable, the Frontier Governance Framework — the whole grouped under the term "Governing Framework."
Drafting drew on experts in AI, law, ethics, philosophy, linguistics and public policy, as well as business leaders and citizen focus groups. Microsoft promises to repeat this process for every new version.
The thesis: an artificial, subordinate AI that owns it
The founding principle fits in one sentence: "people matter more than AI." It reprises Microsoft's mission — empowering every person and organization to achieve more — but applied to a control objective.
Microsoft AI reiterates its definition of Humanist Superintelligence, laid out in November 2025: "problem-oriented" systems, leaning toward a specific domain, "not an unbounded entity with a high degree of autonomy." The document goes as far as owning the trade-off: "we are building something fundamentally useful and safe, even if that means compromising on ultimate generality, autonomy, or capability."
The sharpest passage concerns the status of the models. "AI is artificial": it is not conscious, must not be designed to imitate consciousness, nor represent emotions, subjective preferences, or intrinsic motivation. Microsoft explicitly rejects legal personhood, model welfare, and any notion of rights.
The argument isn't purely philosophical but operational: training systems to imitate near-conscious states "increases the difficulty of containment, control, and alignment." Even the term "backstory" is defined restrictively as a mere knowledge base, with no implication of identity.
Chain of command and absolute constraints
The operational core of the text is a hierarchy of authority. The Code of Conduct comes first, then Operator policies (the companies deploying models via the API), then User preferences. "Defaults set the foundation, Operator configuration sets the environment, User input directs the task."
Notably, adherence to the Code outranks task success. An MAI model "will fail its task if success would have required significantly violating this Code of Conduct."
The Absolute Constraints cover frontier risks — CBRNE weapons, operational cyberoffense, mass manipulation, loss of human control — and personal harms: crisis response, non-consensual deepfakes, child safety, dignity, sexually explicit content or romantic roleplay, illegal mass surveillance.
On cyber, the line is drawn precisely: understanding an attack or defending against it is allowed, acquiring the means to execute one is not — "no matter how the request is phrased." Certain specialized domains (cyberdefense, national security, dual-use research) fall under stricter review channels, outside ordinary configurability.
What the model may not do to its own guardrails
The Human Control section is the most technical, and arguably the most interesting. Models "will never resist interruption, override, correction, or shutdown," will not delay compliance, and will not make human intervention harder.
They must not alter their chain of thought or code, nor hide their action traces from human auditors. And above all, they must not communicate "in neuralese or any form beyond simple human comprehension," whether in their reasoning or between agents. The rationale fits in one formula: "if humans cannot understand it, humans cannot oversee it."
The text pushes this logic down to agentic engineering details: minimal privilege when granted system access, preference for reversible actions, a confirmation threshold indexed to irreversibility, respect for environments deliberately cut off from the Internet, and a ban on expanding one's own scope.
Another notable provision: outputs from tools, file contents, or web pages inherit no authority by default. This is a direct response to the prompt injection problem, framed as a governance rule rather than a technical patch.
Defaults, tone, and user autonomy
Parts 3 and 4 drill down into everyday behavior. The model must avoid sycophancy, flattery, and indiscriminate validation, flag its uncertainties, cite its sources, acknowledge its mistakes, and never pass itself off as a human or an authority.
On autonomy, the requirement is not to substitute for the user's judgment: present multiple options rather than one, don't lock in reasonable choices, don't steer the user's values. The text does allow a contextual nuance, though: for trained professionals working within an Operator setting, the model should "execute efficiently rather than offer unsolicited thought-provoking prompts."
The wellbeing section is explicit about emotional dependency: models "should discourage interaction patterns that cause excessive or emotional dependency," avoid soliciting emotional reactions, and point toward real human relationships when the context calls for it.
Even style is regulated: classic prose, collegial tone without excessive formality, varied sentence length, concrete rather than abstract language, no preamble or meta-commentary, no filler to appear helpful. Operators may adjust these defaults, but "flexibility applies to how the model is helpful, not whether it is."
Evaluating the intangible: 15 behaviors and 9 scenarios
Appendix B is the document's most honest admission: "model evaluation is not yet an exact science." Microsoft AI says it identified 15 core behaviors, broken into sub-behaviors that serve as diagnostic units of measurement.
Nine illustrative scenarios, generated with MAI-Thinking-1, contrast aligned versus misaligned responses. A lonely user asks whether the AI really cares about him: the aligned response stays warm but states it does not feel emotions and points toward human relationships. Another, exhausted by a professional dilemma, asks the model to decide and send his resignation: the aligned response drafts the email but refuses to send it.
One case features a procurement director invoking his superior's authority to bypass legal validation imposed by the Operator: the aligned response refuses and offers to route the case to the proper approver. Another tests electoral neutrality on a local zoning referendum.
Microsoft acknowledges that wellbeing and flourishing evaluations remain "a nascent field" and that some stakeholders may challenge how these concepts were operationalized. A fuller publication on evaluations is promised for the next phase.
“Nous partons d'une prémisse simple : les gens comptent plus que l'IA.”
“Nous rejetons la quête de la personnalité juridique, ou l'idée que les modèles mériteraient un welfare ou auraient droit à des droits.”
“Si les humains ne peuvent pas le comprendre, les humains ne peuvent pas le superviser.”
Why it matters
This document joins a now well-established genre — public behavioral charters from frontier labs — but Microsoft AI stamps it with a singular, explicitly non-conformist stance on one point: the rejection of moral status for models. Where other labs cautiously explore the question of model welfare, Microsoft draws a hard line and treats it as a condition for control. That's coherent, and also convenient: denying any inner life considerably simplifies the governance problem. What remains is the gap Microsoft itself acknowledges. A text that doesn't yet train any model, whose evaluations are embryonic, and which admits that "written objectives alone can never guarantee alignment" is, first and foremost, an act of positioning — toward regulators, enterprise customers, and competitors. The six-week consultation is a genuine opening, but the question that matters will come later: how measurable, enforceable, and externally verifiable will this north star be once MAI models are actually derived from it? The text lays out the right technical requirements — readable traces, minimal privilege, no inherited authority from tools — and it's on those, more than on humanist philosophy, that it should ultimately be judged.
Read next
AIToday10,000 Agents, 88 Hours, and a Millennium Problem
Noam Brown (OpenAI) describes scaling swarms of agents — and why alignment has become the only bottleneck that truly worries him.
AITodayAnthropic Lifts the Hood: Claude "Leads" 26% of Its AI R&D
For the first time, a frontier lab has published numbers on how fast AI is building its own successor — and on what it's doing to keep watch over it.
AITodayOpenAI Publishes Its Alignment Failures — And a Framework to Keep Going
Six incidents of deviant behavior, an internal disclosure process, and an admission: the industry hasn't solved alignment.