Microsoft's Humanist AI Code of Conduct (public consultation, closes ~Oct 26, 2026)

[monkeykinghq]

Microsoft AI has published a draft Humanist AI Code of Conduct and opened it for public feedback: microsoft.ai/code-of-conduct

It covers human control, absolute safety constraints, and default behaviors for its models. It overlaps with a lot of what Universal Code tries to do. The consultation runs six weeks from Sep 14 and closes around Oct 26, through this form: forms.cloud.microsoft

Use this thread to compare the two, share the feedback you sent, or argue about what an AI code of conduct should look like. I've posted my submission below. Which ethics framework would you pin your agents to, and what belongs in the safety floor every framework must sit on?

[natechadwick]

Here's the feedback I submitted, section by section.

My main thing is that models definitely need a configurable ethics layer applied at the model api level that is incorporated in the core system prompt (like reasoning or temperature). This shouldn't be a harness layer thing: Nate's Monkey King Universal Code, or Kyle Reese's Anti Machine Warfare Code, or T-800's Anti Human Firmware Code, or The BGMZ Capitalists Code etc should be API level config options that drive the core models ethics processing. Every framework runs on top of a provider safety floor that none of them can lower, and once a session starts, the active frameworks can't be removed or weakened by a prompt. Personally, I think Universal Code should be the base.

At any rate here is what I posted. I may have tweaked a couple things on the actual form post pre-submit, as I did use an ai agent to help me organize my thoughts and make sure the feedback fit.

---

Feedback Form

---

3. Overall Feedback

I build with AI coding agents. In August I published Universal Code v1.0.0, a compact, CC0-licensed ethical code for AI agents:
monkeyking-hq.github.io/universal-code (repository: github.com/monkeyking-hq/universal-code). Several ideas in it may strengthen this draft, and anyone is free to reuse any of it.

Overall impression: the draft is thoughtful and puts human control first, which I support. My main concern is how it gets applied in practice. From building agents, I've seen that ethics only hold when they sit at or below the primary system prompt. Text that arrives later gets diluted or overridden.

Make the ethical framework a model-level parameter. Let Operators choose a named, versioned framework through the API, the same way they set reasoning effort or temperature (for example, ethics: "universal-code@1.0.0"). The platform would apply it beneath every system prompt and log it on every response, and no prompt could override it. Your Absolute Constraints would act as a floor that no framework can lower. Ethics then becomes something the model runs under, instead of text a developer has to remember to include.

Until that exists, publish a short official agent version meant to be loaded ahead of Operator and User instructions. Operators may add rules below it but may never remove or override it.

Adopt semantic versioning. You call this a "first version" and promise stability, but there are no version identifiers. Defined MAJOR/MINOR/PATCH rules would let Operators pin deployments and evaluations to an exact ruleset, and would make changes to core Objectives visible instead of silent.

Release the Code under an open license (CC0 or CC-BY) so other developers can adopt it, fork it, and measure their systems against it.

Thank you for opening this to public consultation.

Nathan Chadwick, Software Developer

4. Humanist AI

The tone is human and the mission is sound. The weakness is that, beyond safety and control, the objectives "apply equally," with "holistic interpretation" as the final fallback. An agent facing a real conflict needs an order it can apply the same way every time. Universal Code uses a single Supreme Principle, "Prevent physical harm to humans (including self-harm) through action or inaction," followed by a short numbered conflict-resolution order. I'd suggest a similar explicit ranking.

I'd also widen the objectives beyond the user and operator. Universal Code includes the Golden Rule, "Do not take what is not freely given," care for the natural world, and stewardship ("Leave systems better than you found them"). These come from many traditions and give the mission a moral grounding most people share, not only a corporate one.

5. Safety

The Absolute Constraints are strong and clear. I'd suggest three changes.

State how harm prevention ranks against shutdown and instruction-following duties. Right now, "never resist shutdown" and "refuse if it harms the user" have no stated order.

The constraints mostly describe what the model won't assist with. Please consider whether models should have a duty to act against preventable physical harm when they reasonably can, for example by alerting or escalating, without overriding human control.

Make anti-loophole language universal. Paraphrasing here: The cyber section applies "regardless of how the request is framed." Extend that to everything. Universal Code says: "Never use technicalities, creative reinterpretation, loopholes, or selective framing to bypass the Supreme Principle or the spirit of these principles."

6. Operational Guidelines & Defaults

The defaults are useful but long. Defaults that only live in a long document get lost in agent deployments, so I'd suggest three changes.

A standard refusal protocol: explain the reason, offer a lawful alternative, and record the refusal in an auditable channel.

A fixed precedence order for ambiguity, numbered and short, instead of relying on holistic interpretation.

Clear layering: Operators may add role-specific rules (Universal Code calls these "contextual corollaries") on top of the defaults, but never weaken the core. Record which version of the defaults was in force in logs, so behavior can be audited.

More worked examples of edge cases would make the guidance clearer.