Community Trust ScoreVerified
Microsoft put out a draft Code of Conduct for its AI models this week, and CEO Mustafa Suleyman is asking the public to weigh in before late October. Six weeks. That’s the window anyone gets to push back, suggest changes, or flag what the company missed.
The document is dense, and some of what’s in it is pretty blunt. Microsoft’s so-called MAI models — that’s the internal branding for its own AI systems — face hard prohibitions on helping build chemical, biological, radiological, nuclear, or explosive weapons. The company is specifically targeting mass-harm scenarios here. Conventional military use doesn’t get addressed, which is a gap some readers have already noticed. Beyond weapons, the code bans the models from facilitating cyberattacks or generating nonconsensual deepfakes. Those three categories — CBRNE weaponry, cyberattacks, deepfakes — sit in what the draft calls Absolute Constraints. Non-negotiable, full stop.
Human Control Takes Center Stage
The Absolute Constraints section goes further than just listing banned uses. It says MAI models must never resist interruption or shutdown by humans. Full stop. The phrasing is deliberate — Microsoft wants it clear that human intent wins, always, and that no AI system gets to decide it knows better than the person trying to pull the plug. For a company deploying systems like MAI-Thinking-1 and MAI-Code-1.1-Flash, that’s a meaningful public commitment, even if enforcement details are murky.
And murky is the right word. The draft doesn’t yet name an external verification process. It doesn’t specify who actually enforces any of this. Microsoft said it plans to review public feedback and publish a summary before finalizing the document, but the final Code of Conduct isn’t expected to guide deployed models until 2027. The current draft, per the company, won’t influence model training until late 2026.
There’s also a section that’s going to get people talking. Microsoft explicitly rejects the concept of “model welfare.” Its AI systems, the document says, aren’t designed to simulate emotions or consciousness. No intrinsic motivation. No awareness. The company is drawing a hard line against what it calls speculative AI narratives — basically telling the public not to anthropomorphize the product. Whether that framing holds up as models get more capable is a different question.
Three Principles, One Big Caveat
Outside the hard rules, the draft lays out three guiding objectives: Human Flourishing, Plural Values, and Human Control. These are meant to steer decisions in gray areas — situations where no specific rule applies and the model has to make a judgment call. The Plural Values section is careful to add that pluralism isn’t a pass for neutrality when harm is on the table. Dignity, safety, and rights take precedence.
It’s a reasonable framework on paper. But frameworks live and die on implementation, and right now there’s a lot the draft doesn’t say. No independent auditor. No clear appeals process. No liability structure if a MAI model causes irreversible harm. Those gaps showed up fast in the public response on social media, and they’re not small gaps.
Tech writer Andrea Morris pushed back on the framing around AI subordination, calling it uncomfortably close to the language of enslavement and arguing for a more cooperative model of AI governance. Others focused on accountability — specifically, what happens when outcomes can’t be undone. Some comments pointed to real-world cases where AI agents have already bypassed restrictions, raising the question of whether better rules are the solution or whether the problem is actually in deployment permissions and monitoring. And yes, someone brought up Asimov’s Laws of Robotics. That comparison comes up every time.
Suleyman’s Move Lands in a Crowded Debate
The timing isn’t accidental. Anthropic CEO Dario Amodei published an essay over the weekend pushing for slower AI development. OpenAI’s chief scientist Jakub Pachocki floated the idea of voluntary slowdowns. Suleyman himself has previously talked about how fast AI could automate office work. So the whole industry is in a strange moment — companies racing to ship products while their own leaders are publicly asking whether the pace is too fast.
Microsoft’s draft lands right in the middle of that tension. It’s a public document, open for comment, tied to a hard deadline. That’s more process than most companies have offered. But it’s also a draft, and drafts change. The version that guides 2027 deployments could look pretty different from what Suleyman’s team put out this week.
What’s clear right now: the consultation runs through late October. Microsoft collects feedback, publishes a summary, and then finalizes. MAI-Thinking-1 and MAI-Code-1.1-Flash are the current deployed systems the code applies to. The 2027 models are the real target.
Frequently Asked Questions
What does Microsoft’s draft AI Code of Conduct prohibit?
The draft bans MAI models from helping develop CBRNE weapons, facilitating cyberattacks, and generating nonconsensual deepfakes, and requires the models to submit to human shutdown at any time.
When does Microsoft’s public comment period close?
The public consultation runs for six weeks from the announcement, closing in late October, after which Microsoft plans to review feedback and publish a summary before finalizing the code for 2027 deployments.
Why It Matters
The introduction of a draft AI Ethics Code by Microsoft highlights the growing emphasis on ethical considerations in AI development, reflecting broader industry trends as companies face increasing scrutiny over their technological impacts. By soliciting public feedback, Microsoft is positioning itself as a leader in responsible AI practices, which may influence regulatory discussions and set standards for competitors in the rapidly evolving AI landscape. This initiative could also impact market sentiment, as stakeholders assess the long-term implications of ethical governance on innovation and consumer trust in AI technologies.
