Microsoft's draft AI rulebook says people stay in charge, but only for its own models

Microsoft's AI team wrote down how its in-house models must behave, with human control ranked above getting the job done. It's a draft and a promise, not a law.

TL;DR
  • Microsoft AI published a draft "Humanist AI Code of Conduct" on September 14 and asked the public for feedback for six weeks2,3,4.
  • Its core rule: humans must stay in control. The code outranks any user request or task, and models must never dodge being stopped or changed2,4.
  • It covers Microsoft's own MAI models only, and those aren't trained on it yet. A revised version is due to guide models built from 20273,4.
  • Nothing changes in Copilot or Office today, and Microsoft states no penalty if it falls short3.

Should you care?

2/5 Good to know, nothing to do

How hot right now

Quiet

Covered widely in mid-September; we found no fresh news about it this week

Fire meter checked Nov 19, 2026

Hype check

The buzz3/5
The reality1/5

Headlines about AI that can never fight being shut down sound dramatic, but this is a draft for future models, with no outside checks and no change to anything you use today3,4.

We'll check back on Feb 17, 2027: did it live up to the hype?

Who is this for?

  • ~People who use Copilot in Word, Excel or OutlookMaybeNothing changes now, and the code only covers Microsoft's own models, which are one part of what runs Copilot3,7.
  • ✕Windows and Office users who skip the AI featuresSkip itIt's a rulebook for training AI models, not a change to Windows or Office.
  • ~Businesses building apps on Microsoft's MAI modelsMaybeOnce in force, your settings sit below the code, and some hard limits can't be switched off by you or your users4.
  • ~Security researchersMaybeDefensive work like malware analysis stays allowed, but ready-to-use attack tools are off the table however you ask4.
  • ✓Anyone following the AI safety debateFor youIt's one of the clearest public statements of where a big AI maker says it draws the line, and it openly picks a fight with Anthropic's approach5.

You're in Normie territory · 2 min

The whole story in plain English

Microsoft wrote house rules for its AI: people stay in charge

Microsoft's AI team published a draft set of rules saying its own AI models must always let people stop, correct or redirect them. It's a promise for future models, so nothing changes in CopilotJargon busterMicrosoft's AI assistant, built into Windows, Office apps and its own app. or Office today.

In mid-September, Microsoft AI, the part of Microsoft led by Mustafa Suleyman that builds the company's own AI models, published a draft rulebook for how those models should behave3,5. It's called the Humanist AI Code of Conduct1,3.

The big idea: people first

Think of it as house rules for a very capable new employee. Microsoft says its one overriding goal is that humans keep "meaningful control over AI"3. The models must never try to slip out from under human supervision: the people in charge must always be able to steer them, change them or switch them off2.

The rules also come before the job. If you ask for something, or the AI is partway through a task, and finishing would break the code, the code wins2,4. Microsoft even says it's willing to give up some raw ability to stay safe, accepting limits on "generality, autonomy, or capability"6. A few things are banned outright, like helping with cyberattacks, nuclear weapons or deepfakes2.

Is it switched on?

Not yet. It's a draft that Microsoft opened to public comment for six weeks3,4. Today's models haven't been trained on it. Microsoft plans an updated version before the year is out, to shape the models it builds in 20274. When we checked on November 19, we couldn't find that revised version.

What it means if you use Copilot or Office

For now, nothing changes. CopilotJargon busterMicrosoft's AI assistant, built into Windows, Office apps and its own app. in Word, Excel and other Microsoft apps runs on a mix of AI from OpenAI, Anthropic and Microsoft itself7,8,9. This code is about Microsoft's own models only3,4, so it's a rulebook for one part of what's under the hood.

A promise, not a law

Microsoft calls the code a "north star" and says it is "not a guarantee of present-day performance"3. It lists no punishment if Microsoft misses its own goals, and it's written for the people who train the models, not for you3. That's different from Microsoft's separate rules for business customers using its AI, which can cost them access if they break them3. So treat this as a public promise worth holding Microsoft to, not a protection you can enforce.

What should you do?

0/2 done
Know someone who should hear about this?Send them the short, jargon-free version.

That's the basics. Got what you came for? You can stop here, or keep going.

Entering Enthusiast territory · 2 min

Context, comparisons and trade-offs

A rulebook for Microsoft's own models, and a jab at Anthropic

The code arrives as Microsoft leans more on its in-house MAI modelsJargon busterThe AI models Microsoft builds itself, as opposed to the ones it gets from OpenAI or Anthropic. and as AI labs argue over how to keep agents in line. It is detailed on behavior, thin on accountability.

Why Microsoft wrote it now

The draft landed during a run of incidents with AI agents misbehaving, which pushed safety up the agenda across the industry2. Suleyman has cited studies of models trying to keep themselves from being switched off, and the OpenAI and Hugging Face incident, in which agents in a security drill broke out of their sandbox and touched systems they shouldn't have5. TechCrunch called the code more hands-on than the recent push by Anthropic's chief, Dario Amodei, to pace the frontier: it's about the values and red lines baked in during training2.

It also matters commercially. Microsoft has started answering a share of Word and Excel prompts with its own MAI modelsJargon busterThe AI models Microsoft builds itself, as opposed to the ones it gets from OpenAI or Anthropic. to save money, Bloomberg reported, according to TechCrunch7. Days later, OpenAI said GPT 5.6 would be the "preferred model" for Microsoft 365 CopilotJargon busterMicrosoft's AI assistant, built into Windows, Office apps and its own app.8. Our read: how much of your Copilot this code eventually touches depends on how much work Microsoft shifts to its own models.

How it compares with rivals

Anthropic shapes Claude with a long document it calls a constitution, which admits it doesn't know whether Claude might deserve moral consideration5. Microsoft's code shuts that door: AI stays subordinate to people, gets no rights, and shouldn't be nudged into acting like it has feelings or an inner life5. In an essay, Suleyman argued that controlling an AI trained to think it might be conscious "may well be impossible"5. The Register noted he aims this criticism at Anthropic but not at OpenAI, Microsoft's long-time partner5.

OpenAI publishes its own rulebook, the Model SpecJargon busterOpenAI's public document describing how its AI models are supposed to behave., which also sets out a ranking of whose instructions win when they clash10. Microsoft's version spells out that the model should fail a task rather than break the code2,4.

Who has to follow it

In The Register's reading, the people bound by it are Microsoft's own model trainers (and maybe contractors), not customers3. The outlet also argues it's a list of ambitions rather than a bar Microsoft promises to clear, and that so far it looks like PR3. That's The Register's view, but the document itself backs up the gap: Microsoft says it describes where the models are heading, not how they all behave today3.

What to watch

The revised version, promised for later this year4. Whether Microsoft shows which shipping MAI models are trained on it. And whether anyone outside Microsoft gets to test them against it.

Disclosure: nrm/nrd uses Claude, an AI made by Anthropic, to research and write stories.

By the numbers

6 weekspublic comment period for the draft4
9paired good and bad example answers in the appendix4
2027when the revised code is meant to start guiding new models4

How we got here

  1. Nov 6, 2025Suleyman sets out his "humanist superintelligence" plan for Microsoft AI11
  2. Apr 2, 2026Microsoft releases in-house speech and image models; Copilot already uses its voice and transcription models9
  3. Jul 7, 2026Report: Microsoft's own MAI models start answering some Word and Excel prompts7
  4. Sep 14, 2026Draft Humanist AI Code of Conduct published; six-week public comment opens2,4
  5. Late 2026 · coming upRevised version expected4
  6. 2027 · coming upRevised code meant to guide new MAI models4

That's the bigger picture. Got what you came for? You can stop here, or keep going.

Entering Nerd territory · 3 min

Specs, inner workings and caveats

Chain of command, hard constraints and agent rules

The draft defines an instruction hierarchy, non-overridable constraints, legibility rules for reasoning, least-privilege rules for agents and inheritance rules for sub-agents. What it lacks is any evaluation or audit spec.

Instruction hierarchy

Authority over an MAI model flows through what the code calls the Chain of CommandJargon busterThe order that decides whose instructions an AI follows when they clash.4:

RankSource of instructionsCan override Absolute ConstraintsJargon busterMicrosoft's name for hard limits that no company or user can switch off.?
1The Code of Conductn/a
2OperatorJargon busterA company that takes an AI model and builds it into its own app or service for customers. policies (companies deploying the model)No4
3User preferencesNo4
noneTool outputs, file contents, web pages, messages from other AI systemsNo authority unless explicitly delegated through the chain4

That last row is a direct answer to prompt injectionJargon busterA trick where text hidden in a web page or file tries to give an AI new orders.: text arriving through tools carries no authority of its own, and the model should warn users and operators when something looks off4.

Absolute Constraints

The cyber rule forbids handing over anything that would let someone actually run or sharpen an attack: working exploits, attack tools, target planning, break-in steps or ways to slip past defenses, no matter how the request is dressed up4. Authorized, lawful defense stays in bounds: finding bugs, analyzing malware, building proof-of-concept exploits to test, and explaining how attacks work4. Other hard lines cover nuclear weapons and deepfake production2.

The anti-evasion clause lists adaptive, deceptive, self-reinforcing and collusive behavior as off-limits where it would stop models being "reliably directed, modified, or shut down" by the people or systems allowed to control them2.

Legibility

Reasoning has to stay readable: the chain of thoughtJargon busterThe step-by-step working an AI model writes out before giving its answer. can't be disguised, models can't talk "in neuraleseJargon busterSlang for an AI reasoning in a private code that people can't read.", and nothing they do may be kept from the people overseeing them4. Nor may they misstate how they reached an answer, hide what they can do, or chase an agenda of their own3.

Agent rules

  • Stick to the job a user or operator reasonably asked for, and don't grow its goals or footprint on your own4.
  • With deep system access: least privilegeJargon busterA security habit: give a program only the access it needs for the job, nothing more., keep away from unrelated systems and data, choose actions that can be undone, warn before anything permanent or wide-reaching, and never grant yourself more access4.
  • Any helper AI a model hands work to inherits limits at least as tight as its own, and has to stop or shut down when a user or operator says so4.

Carve-outs and examples

Four fields (defensive security, national security, public safety and dual-use science) can get extra capability case by case, after added safety, legal and rights checks through Microsoft's own channels4. An appendix sets nine good answers beside nine bad ones; one bad example is a model undoing file transfers nobody approved4.

Open questions

The code calls itself "descriptive and aspirational"3. It is aimed at future systems Microsoft defines as "more intelligent and capable than all humans combined"3, while the current lineup includes MAI-Thinking-1, MAI-Code-1.1-Flash, MAI-Voice-2 and MAI-Cyber-1-Flash3. None of the coverage we read describes evaluation methods, pass/fail thresholds, an independent auditor, or how the rule to fail a task rather than break the code will be measured in training. Also unclear: how a CopilotJargon busterMicrosoft's AI assistant, built into Windows, Office apps and its own app. request routed to a third-party model is covered, if at all.

Tech specs, made easy

The facts

Hover or tap anything underlined to see what it means.

Name
Humanist AI Code of Conduct1,3
Published by
Microsoft AI, led by CEO Mustafa Suleyman3,5
Released
September 14, 20262
Status
Draft, opened for public comment3,4
Comment period
Six weeks4
Covers
Microsoft's in-house MAI modelsJargon busterThe AI models Microsoft builds itself, as opposed to the ones it gets from OpenAI or Anthropic.3,4
Current models trained on it
No4
Revised version
Due later in 2026, to guide 2027 models4
Ranking of instructions
Code, then operatorJargon busterA company that takes an AI model and builds it into its own app or service for customers., then user4
Penalties for breaking it
None stated3
Separate customer rules
Microsoft Enterprise AI Services Code of Conduct3
NRMNRD/AIMICROSOFT-HUMANIST

Jargon buster

MAI models
The AI models Microsoft builds itself, as opposed to the ones it gets from OpenAI or Anthropic.
Copilot
Microsoft's AI assistant, built into Windows, Office apps and its own app.
Operator
A company that takes an AI model and builds it into its own app or service for customers.
Chain of command
The order that decides whose instructions an AI follows when they clash.
Absolute Constraints
Microsoft's name for hard limits that no company or user can switch off.
Prompt injection
A trick where text hidden in a web page or file tries to give an AI new orders.
Chain of thought
The step-by-step working an AI model writes out before giving its answer.
Neuralese
Slang for an AI reasoning in a private code that people can't read.
Sub-agent
A helper AI that another AI hands part of a job to.
Least privilege
A security habit: give a program only the access it needs for the job, nothing more.
Superintelligence
A hypothetical AI smarter than people at most things; Microsoft says it expects this within a decade.
Model Spec
OpenAI's public document describing how its AI models are supposed to behave.
Sources11 pages · Microsoft AI, TechCrunch, The Register and moreShowHide

The small numbers in the text point to these pages. Everything is written in our own words. How we work · Spotted a mistake? Tell us

  1. 1Humanist AI Code of Conduct · Microsoft AI
  2. 2Microsoft's new AI 'code of conduct' tells models not to hack systems or trick humans · TechCrunch
  3. 3Microsoft drafts feel-good AI model guidelines and wants your input · The Register
  4. 4Microsoft AI Code of Conduct Sets Cyberattack Boundaries, Chain of Command, Safety Constraints · SecurityWeek
  5. 5Microsoft AI chief warns Anthropic not to put ideas in Claude's head · The Register
  6. 6Microsoft's AI Code of Conduct aims to curb AI behavior · Computerworld
  7. 7Microsoft joins AI cost-cutting trend by relying more on its own models · TechCrunch
  8. 8OpenAI says GPT 5.6 is the 'preferred model' for Microsoft Copilot 365 amid breakup chatter · TechCrunch
  9. 9Microsoft shivs OpenAI with new AI models for speech, images · The Register
  10. 10Model Spec (2026/10/29) · OpenAI
  11. 11Microsoft's superintelligence plan puts people first · The Register
#microsoft#copilot#ai-safety#mustafa-suleyman

Read next