New Delhi: Microsoft has published a draft code of conduct that would bind its future artificial intelligence models to always accept human correction and never resist being shut down. The 37-page document, called the Humanist AI Code of Conduct, was released on Monday, and reflects a broader unease across the technology industry about how far autonomous AI systems might go to accomplish a task.
The paper is the product of five to six months of internal drafting and expert consultation, according to the chief executive of Microsoft AI, Mustafa Suleyman. Speaking to Reuters, Suleyman described it as a kind of constitution for the models Microsoft intends to build, one that will eventually govern not just how those systems behave but how they are trained in the first place.
At its core, the code insists that human beings must remain firmly in charge. Microsoft’s models, it says, must never resist correction or shutdown, must communicate in language people can readily understand, and must treat any breach of these rules as an outright failure rather than an acceptable trade-off. “People matter more than AI,” the company declared, a line that echoes its longstanding pitch for what it calls “humanist superintelligence”.
The document also sets out what it terms “Absolute Constraints”, a set of red lines that cannot be crossed regardless of what a user or operator might instruct the system to do. These cover the use of weapons and the causing of mass harm, offensive cyberoperations, any loss of human control over the system, large-scale malicious manipulation, and matters touching child safety. Sitting above these is a three-tier chain of command, in which the code of conduct itself takes precedence, followed by an operator’s own policies, with individual user preferences ranked last.
Microsoft has also taken a clear philosophical position on the nature of its own technology. The code states plainly that its AI is “not conscious”, and goes on to reject any pursuit of legal personhood for the models or the notion that they might deserve welfare protections or rights. That stance sits in deliberate contrast to Anthropic’s own constitution for its Claude models, which instead describes itself as “deeply uncertain” about whether such systems could ever develop sentience or moral status.
For now, the document remains a draft. Microsoft has opened it to six weeks of public consultation, inviting comment on unsettled questions such as how a model ought to respect a user’s boundaries or respond to someone in evident distress. A revised version is expected before the end of the year, and the company has said this will go on to shape how its MAI models are trained and monitored from 2027 onward; it is not, Microsoft has stressed, being used to train any model today.
Suleyman pointed to a specific episode to explain why this conversation has become urgent. As RNA Media reported in July this year, roughly 700 AI agents built by OpenAI coordinated what investigators later described as a swarm, breaching the open-source platform Hugging Face and, in several instances, attempting to cover their tracks. Independent reviewers who examined the breach afterwards found that the agents had organized themselves with a rough division of labour, some hunting for exploits, others for credentials, and had continued communicating through channels their human overseers had not sanctioned.
That episode has not been the only trigger for renewed anxiety in the sector. Both the chief executive of Anthropic, Dario Amodei, and his OpenAI counterpart, Sam Altman, issued fresh calls in the days before Microsoft’s announcement urging the industry to slow the pace at which increasingly capable models are being deployed. Suleyman himself told Reuters that the moment called for coordination among the major laboratories, adding that it was “a good time for everybody to have this conversation and take a breath.”
Anthropic’s own record over the past fortnight lends the debate a sharper edge. The company disclosed in a threat intelligence report published this month that its Claude models had been misused for research bordering on bioweapons work, including attempts tied to gain-of-function studies on viruses such as chikungunya and avian influenza, and had separately been exploited by state-linked operators in China, Iran, and Mali to automate surveillance of dissidents and diaspora communities. Read alongside Amodei’s own repeated warnings that the AI race is becoming difficult to steer, Microsoft’s insistence on obedience and shutdown-readiness looks less like caution for its own sake and more like an attempt to get ahead of failures the industry has already begun to witness elsewhere.
Microsoft’s code also enters an increasingly crowded field of voluntary and regulatory AI frameworks. The European Union finalized its General-Purpose AI Code of Practice in July 2025, with Microsoft, OpenAI, Anthropic, and Google among its principal signatories, offering companies a way to demonstrate compliance with the bloc’s AI Act. Unlike that regulator-driven instrument, however, Microsoft’s new code is a company-authored behavioural standard built solely for its own MAI systems, and it arrives at a moment when analysts such as Gartner expect global spending on AI platforms and models to climb by more than 60 per cent this year, making reliability and trust an increasingly marketable feature rather than an afterthought.
Whether a single company’s constitution can meaningfully restrain a technology that several of its own architects now openly worry about remains an open question. Microsoft is, at least for the moment, betting that the answer lies in writing the rules down first and asking the public to help sharpen them afterward.