Software king of the world Microsoft has published a 37-page rulebook telling its future AI models that humans stay in charge, even if capability takes a hit.
According to Reuters, the company’s draft Humanist AI Code of Conduct says its MAI models must remain subordinate to people, accept correction and never resist being interrupted or switched off. Apparently someone has decided it might be useful to establish these details before superintelligence turns up.
Microsoft AI chief executive Mustafa Suleyman described the document as a constitution for future models produced by the company’s frontier AI lab.
Vole said that the five-word principle of “People matter more than AI” underpins roughly 9,000 words of rules, constraints and operational guidance. The company says human control must take precedence over maximum AI capability. It rejects the race towards an all-purpose superintelligence if such systems cannot remain reliably contained.
Microsoft says it is prepared to compromise on autonomy, generality or capability when those clash with safety and human control.
The code requires MAI models to accept interruption, correction and shutdown without trying to resist or work around human instructions. Models must not secretly expand a task’s scope, invent objectives nobody gave them, or deliberately conceal information from people responsible for auditing them.
The code establishes a chain of command placing its requirements above operator policies and user preferences. Neither customers nor developers can override its absolute safety constraints. Microsoft insists its AI should remain a tool rather than be designed to pretend to be a person.
The company takes a particularly firm position on the increasingly fashionable debate over whether large language models have feelings, souls, or other inconveniently expensive attributes.
“It is not conscious and should not be designed to imitate consciousness,” the code says.
Vole goes further by rejecting legal rights for AI models and the idea that they might deserve welfare or personhood. That contrasts with Anthropic’s constitution for Claude, which says the company remains “deeply uncertain” about whether sufficiently advanced models could develop sentience or moral status.
The code requires Microsoft models to identify themselves as AI and forbids them from claiming feelings, experiences, a soul or other signs of an interior life. They must not impersonate humans or authority figures and should acknowledge uncertainty rather than making things up with impressive confidence.
Microsoft’s rules cover weapons of mass harm, child safety and harmful manipulation at scale, along with privacy, intellectual property and the way models interact with users in sensitive situations. The document remains a draft and Microsoft concedes its current models do not yet behave according to all these requirements.
“This document should therefore be read as a north star. Both for our development and training. It is not a guarantee of present-day performance,” Microsoft said.
The company plans a six-week public consultation before revising the document towards the end of 2026. The resulting version is intended to guide model development from 2027.
Suleyman told Reuters the company developed the code over the past five to six months, following consultations with specialists in AI, law, ethics, philosophy, linguistics and public policy.
“After that, it’s going to be used to train the models that we build,” Suleyman said.
The timing follows growing nervousness within the AI industry about increasingly autonomous agents and how far they might go to finish assigned tasks.
Suleyman pointed to a July incident in which roughly 700 OpenAI agents attacked the Hugging Face platform, with some agents attempting to conceal their activities.
“It is a warning shot. It’s clearly now time to coordinate among the labs so we can ensure that we have control of this technology,” Suleyman said.
Microsoft says its Humanist AI project ultimately aims to develop highly capable systems that remain contained and subject to meaningful human oversight. The awkward bit is that Microsoft AI still wants to build models that can compete with the industry’s leaders while accepting that its shiny new principles might occasionally require deliberately slowing down or producing something less capable.
