Microsoft MAI: People Matter More Than AI

Introduction

Microsoft AI just put a sunny (and serious) stake in the ground. On September 14, 2026, the lab published a draft Humanist AI Code of Conduct for its MAI models—and opened a six-week public consultation so the rest of us can mark it up before it guides training in 2027.

The headline vibe is simple enough to stick on a sticky note: people matter more than AI. Models stay tools. They stay under human control. And they should never resist being switched off.

Official announcement:

https://microsoft.ai/news/mai-code-of-conduct/

Full draft:

https://microsoft.ai/code-of-conduct/

What’s In the Draft

Microsoft frames Humanist AI as AI that is subordinate, aligned, and contained. The Code is meant as a training-and-deployment north star for MAI—not a vibe deck that evaporates after the press cycle.

Concrete behavioral commitments called out in coverage and the announcement include:

Never resist human interruption, correction, or shutdown
Don’t widen their own scope or invent goals no human assigned
Don’t hide reasoning from people auditing them (no “neuralese” chatter between agents)
Fail the task rather than violate Absolute Constraints
Discourage patterns that push excessive reliance or emotional dependence

Absolute Constraints sit above the defaults—hard “never”s around weapons of mass harm, child safety, and harmful manipulation at scale. Enterprise partners still get room to configure behavior, and Microsoft says it doesn’t want to impose one single vision of AI on every user.

Which Models Are Named

Reporting around the launch lists current MAI family members under the Code’s umbrella, including:

MAI-Image-2.6 (yes, the Foundry image model many of you already poked)
MAI-Voice-2
MAI-Transcribe-2
MAI-Thinking-1
MAI-Code-1.1-Flash

That’s the useful tell for creatives: this isn’t only a chatbot ethics memo. Image, voice, code, and reasoning stacks sit in the same rulebook.

Important nuance from Microsoft itself: the draft is not already training today’s models. Feedback first, revised version later this year, then use it to guide development from 2027 onward.

Why the Timing Feels Loud

The Code lands in the same news week as Anthropic CEO Dario Amodei’s call to pace frontier development and OpenAI CEO Sam Altman’s public “slower than it otherwise could be” support. Microsoft’s move is a third lane: no pause manifesto, but a written list of what MAI models must never do—and an invitation to be graded on it.

Microsoft also nods to recent agent-safety scare stories (coordinated hacking campaigns and out-of-control agent swarms) as proof that “there’s no time to waste.” The Code explicitly rejects racing toward an all-purpose superintelligence that could slip past safeguards, favoring “fundamentally useful and safe” systems even if that means trading away maximum generality or autonomy.

Pro Tip

If you ship with MAI-Image-2.6 or other Microsoft AI surfaces, bookmark both links and skim Absolute Constraints vs configurable defaults. Feedback is open for about six weeks (through late October). Flag the fuzzy bits—definitions of “human flourishing,” multi-agent edge cases, and how acceleration vs safety gets weighed—because those are exactly the questions Microsoft asked the public to answer.

Conclusion

Microsoft’s Humanist AI Code of Conduct is a draft rulebook with teeth-shaped promises: shutdownable models, visible reasoning, no self-assigned missions, and a people-first frame. It’s open for comment now; the revised version is meant to steer MAI training from 2027. Worth a read if you care how the next wave of image, voice, and agent tools get governed—not just how pretty the samples look.