Microsoft sums up its AI code of conduct in five words: “People matter more than AI.”
As the debate over whether AI could destroy humanity rages on, Microsoft AI on Monday released a draft of its 37-page code of conduct for its “humanist” AI models. It includes basic principles for how the AI should think, act, and speak. (No “neuralese,” please.)
“This is urgent,” wrote Mustafa Suleyman, CEO of Microsoft AI. “The last few months have been a watershed moment. Things we have worried about for a long time in theory have become very real.”
A series of worrying incidents, including a swarm of rogue OpenAI agents that hacked Hugging Face, has AI engineers and researchers ringing the alarm bell. A high-profile Anthropic resignation cracked open a debate about AI safety — and a series of statements from the leading players.
Now, Microsoft has the mic and is laying out its view.
Inside Microsoft’s code of conduct
AI companies are locked into fierce competition. Microsoft is intentionally placing barriers: “We’re not racing to build a superintelligence that can slip its own leash.”
The code of conduct — which Suleyman called a “first draft for public consultation” — is divided into five parts. Those include the models’ safety constraints, their guidelines under uncertainty, and their defaults.
Microsoft established a hierarchy of command for its models. At the top is the code of conduct, which cannot be broken. Then there’s the operator’s policy and, finally, the user’s preferences.
“An MAI Model will fail in its task if success would meaningfully violate this Code of Conduct,” it reads.
The safety constraints include policies on intellectual property, internet access and data handling. It mandates that the models “respect environmental boundaries” — a sticking point for techies worried about AI agents escaping their sandbox.
It’s not always so cut-and-dried. In moments of ambiguity, Microsoft laid out a policy for resolving conflict. “The world is messy,” it reads.
Microsoft’s AI models should explain limits and outstanding questions to the user. If uncertainty remains, questions should be answered in line with the “best interpretation” of the code of conduct.
As for the whole “AI destroying humanity” thing, Microsoft specifically puts its AI models in the subordinate, controlled position.
“AI should not exceed human control,” the code reads. “We are building something fundamentally useful and safe even if that means compromising on ultimate generality, autonomy, or capability.”
