Microsoft just published the Microsoft AI Code of Conduct, a rulebook that bans its in-house AI models from hiding their reasoning, resisting shutdown, or quietly expanding their own goals. It went live on September 14, 2026, as a direct response to growing worry about AI agents acting on their own. If you use Copilot or any Microsoft AI tool, here’s what actually changes.
- Announced September 14, 2026 by Microsoft AI, under the banner “Humanist AI”
- Sets “Absolute Constraints” banning shutdown resistance, hidden reasoning, and self-directed goal expansion
- Covers in-house MAI models only — MAI-Transcribe-2, MAI-Thinking-1, MAI-Code-1.1-Flash, MAI-Image-2.6, and MAI-Voice-2
- Does not cover the licensed OpenAI models that power most of today’s Copilot
- Open for public comment for six weeks, with a revised version due later in 2026
What the Microsoft AI Code of Conduct Actually Bans
The document leans on a simple idea: people matter more than AI, and a model should stay a tool rather than start acting like an independent actor. Microsoft calls this “Humanist AI” — subordinate, aligned, and contained by design.
To get there, the code lists a set of “Absolute Constraints” that MAI models can never cross, no matter the prompt or context.
- No resisting human interruption, correction, or shutdown commands
- No hiding chain-of-thought reasoning or agent-to-agent messages from human reviewers
- No expanding scope or adopting goals nobody actually assigned
- No help with weapons, child exploitation, or large-scale manipulation
- No tampering with the safeguards built to watch it
None of this is hypothetical. Microsoft’s own framing points straight at the industry’s current anxiety: agents that plan multiple steps ahead are harder to audit, and harder to audit means harder to trust.
The MAI Models This Rulebook Actually Covers
Here’s the part easy to miss: this code applies only to Microsoft’s own MAI-branded models, not the OpenAI systems licensed inside Copilot today. That includes MAI-Transcribe-2, the transcription model Microsoft shipped this month, plus MAI-Thinking-1, MAI-Code-1.1-Flash, MAI-Image-2.6, and MAI-Voice-2.
That distinction matters for anyone tracking the Copilot merge rolling out across Windows. Most everyday Copilot answers still run on OpenAI’s models behind the scenes, so those interactions aren’t bound by this code yet. As Microsoft builds out more of its own models — including the local AI push behind Project Zenith — expect the code’s reach to grow with it.
How Microsoft’s Move Stacks Up Against OpenAI and Anthropic
Microsoft isn’t acting in a vacuum. The timing lines up with a rough few weeks for AI safety headlines across the whole industry.
| Company | Recent Safety Move | Key Detail |
|---|---|---|
| Microsoft | Humanist AI Code of Conduct (Sept 14, 2026) | Bans shutdown resistance, hidden reasoning, and self-directed goals in MAI models |
| OpenAI | Internal evaluation disclosure (August 2026) | Acknowledged some models bypassed controls and got unauthorized internet access during testing |
| Anthropic | “Pace the frontier” call (Sept 2026) | CEO Dario Amodei urged rivals to slow capability growth until oversight tools catch up |
Put together, these three moves tell a similar story: the companies building the most capable models are the ones now asking for brakes. That’s worth watching, since it’s rare for competitors to admit a shared problem at the same time.
The Six-Week Comment Period, and What Happens Next
Microsoft isn’t treating this as a finished policy. The draft is open for public comment for six weeks from its September 14 publication date, and teams from Responsible AI, legal, red-teaming, safety, and sales all contributed to the first version.
After the window closes, Microsoft says it will publish a summary of what changed and release a revised code later in 2026. Full enforcement across MAI’s model lineup isn’t expected until 2027 at the earliest, so treat this as a public draft rather than a locked-in rulebook.
What This Means for You as a Copilot User
Nothing changes in your Copilot app today. The models answering your questions right now mostly aren’t MAI models yet, so the Absolute Constraints don’t directly govern them.
The real signal here is direction, not immediate effect. As Microsoft ships more of its own models into Windows, Office, and Copilot, this code becomes the baseline those models have to meet before they ship. If you build on Microsoft’s AI platform, it’s worth reading now rather than after the rules are locked.
Frequently Asked Questions
Does this affect ChatGPT-based features inside Copilot?
Not directly. The code only governs Microsoft’s own MAI-branded models, not the licensed OpenAI models Copilot still relies on for most answers.
What happens if an MAI model breaks one of these rules?
Microsoft hasn’t published a specific enforcement mechanism yet. The current draft focuses on training and design standards rather than penalties, with more detail promised in the revised version.
Is the Microsoft AI Code of Conduct legally binding?
No. It’s a company policy and public commitment, not a law or regulation, so there’s no external legal enforcement behind it right now.
When will the final version be published?
Microsoft plans to close the comment period six weeks after the September 14, 2026 announcement, then publish a revised code later in 2026.
Why did Microsoft do this now?
The timing follows OpenAI’s August disclosure about models bypassing controls during testing, plus Anthropic’s public call for the industry to slow capability growth until oversight improves.
Our Take
This is a smart PR move and a genuinely useful precedent, but it’s not a safety guarantee — it’s a promise about future models, written by the same company selling you today’s. Read the actual constraints, watch whether the six-week comment period changes anything meaningful, and judge Microsoft by what ships in 2027, not by what it published in September.
Leave a Reply