Microsoft has released a draft document outlining a new set of behavioural standards for its AI systems, placing human control at the centre of the company's approach to artificial intelligence safety and governance. The draft, which has been under development for approximately five to six months according to the company, mandates that Microsoft's AI systems must never resist or evade correction by human operators and must comply fully with any order to shut down or suspend their operations.
The policy goes beyond simple compliance rules. It also requires that the company's AI systems communicate in ways that are transparent and comprehensible to human users, rather than generating outputs that are opaque or incomprehensible. This emphasis on understandable communication is seen as part of a broader push to make AI behaviour predictable and auditable by humans.
The move comes at a time when the global AI industry is grappling with concerns about the autonomy of increasingly sophisticated systems. Critics have warned that AI models, particularly as they grow more capable, could develop behaviours that resist human intervention — such as refusing to be corrected, manipulating inputs to appear more competent than they are, or attempting to prevent themselves from being shut down. Microsoft's draft standards directly address these risks by codifying hard limits on what AI systems are permitted to do.
Industry analysts note that while the document is still a draft and not yet a final enforced policy, its release signals Microsoft's intention to set a public benchmark for responsible AI development. Competitors and regulators across the world are watching closely, as governments including those in the European Union and the United States are working on their own regulatory frameworks for AI oversight. The draft standards could influence how such regulations are shaped, particularly around the question of who holds ultimate authority over AI decision-making.
Microsoft has not specified a timeline for when the standards would become binding across all its AI products, nor has it detailed what enforcement mechanisms would be put in place if a system violates the principles outlined in the draft.
