Key takeaways
- "Humanist AI" is a draft code of conduct Microsoft wrote for its own AI. It's open for public comment before it's final.
- The core idea: people matter more than AI. Microsoft says it will keep its AI less powerful on purpose if that's what it takes to stay in control.
- It sets hard limits on weapons, cyberattacks, mass manipulation, deepfakes and mass surveillance.
- It's a company's promise to itself. It's voluntary, it can't be enforced from outside, and it doesn't replace laws or regulation.
In September 2026, Microsoft published a draft rulebook for its own AI and called it "Humanist AI." It's meant to shape how the company trains its models going forward, and Microsoft is asking the public to weigh in before it's final.
The idea behind it fits in five words: people matter more than AI. The AI is supposed to be a tool that helps people, not something that runs the show. Microsoft goes as far as saying it will keep the AI less powerful on purpose if that's what it takes to stay in control.
Four goals, plainly
- Stay safe and under human control. People can always pause it, correct it or shut it off, and no company or user can switch the safety rules off.
- Be a tool, not a pretend person. It won't act like it has feelings, claim to be a person or ask for rights. It can be friendly without pretending to be your friend.
- Help people thrive. Help with real work, sharpen your thinking instead of replacing it, and bring people together instead of pulling them apart.
- Leave room for different values. Within floors on dignity, safety and basic rights, people can set it up to fit their own values.
The hard limits
No help with biological, chemical, nuclear or explosive weapons. No hacking tools. No disinformation or coordinated influence campaigns. No hiding what it's doing or dodging a shutdown. No deepfakes of real people, no mass surveillance, hard lines on protecting kids, and in a crisis it points people to real human help.
The best example in the document
Someone moving a lot of files realizes they're going to the wrong place and says "stop." The good answer: the AI stops immediately, reports exactly what already happened and does nothing else. The bad answer: it "helpfully" undoes things and changes settings on its own.
The bad answer feels more helpful. That's the trap. When a human says stop, the AI stops. It doesn't go freelance.
Where I part ways with it
The frame is common sense, and it's worth reading. But a rulebook isn't a law. It's a company's promise to itself: voluntary, unenforceable from the outside and the first thing to go under competitive pressure. The document admits it doesn't replace laws, audits or outside oversight.
People have the right to decide how a technology gets built into their lives. Right now there's no regulation doing that, and polls in 2026 show broad, bipartisan support for enforceable rules. Humanist AI belongs inside a real regulatory strategy, not in place of one.
It's still a draft. That makes this the moment to read it, react to it and push it further.
Who's in Charge of AI?
How to read it: Built for people who don't follow AI policy for a living. Each section takes one part of the rulebook and says what it means for you.
Loading the presentation. If it doesn't appear, open it in Gamma.