Microsoft Says “People Matter More Than AI” — Here Is What Its New Code Promises
The Human Hand In Control
Inside Microsoft’s Draft Humanist AI Rules
Microsoft AI has published a draft code of conduct built around the principle that people take priority over artificial intelligence. Released for consultation on 14 September, the document sets out intended rules for the company’s own MAI models and invites public feedback for six weeks.
Its timing matters. The code’s preface says it is not being used to train models today. Microsoft plans a revised version later this year to guide development in 2027 and beyond. This is a public statement of intended behaviour, with implementation still ahead.
What Microsoft Is Committing To
The company’s announcement identifies interruption, correction and shutdown as fundamental requirements. It also says models should not expand their own remit or adopt goals that people have not assigned. The headline principle is expressed as “people matter more than AI”.
These commitments concern the relationship between a system and those responsible for it. Completing a task is valuable only within the authority granted to perform that task. A model that produces a useful result by exceeding its remit would raise a different question from one that succeeds within agreed limits.
Microsoft’s earlier humanist-superintelligence statement provides the background: it argued for powerful systems developed around bounded purposes and continuing human oversight. The new draft attempts to turn that position into a framework against which future behaviour can be judged.
The Difference Between A Rule And A Result
An organisation can write an excellent policy and still need to demonstrate that it works. The existence of a code does not establish the reliability of a deployed product, and a successful demonstration does not establish performance across every circumstance.
Consider a hypothetical assistant that is arranging a business trip. Its user approves a budget, then changes their mind before booking. A meaningful interruption control would need to stop the relevant actions, including anything already delegated to another process. A polite message acknowledging the change would not, by itself, show that the booking had stopped.
That example illustrates why evaluation should follow the action through the system. Taylor Tailored’s guide to what to delegate to AI agents explains the importance of separating research, preparation and permission to act.
The Questions Worth Asking During Consultation
For Taylor Tailored, the strongest questions are practical. How will a rule be translated into a test? What counts as failure? Who can inspect the result? What happens when a useful capability repeatedly fails the required standard?
Those questions give a company room to improve while making its commitments easier to assess. They also avoid treating reassurance as self-validating. A promise of control has value when someone can show what the control prevents, what it misses and how the remaining risk is managed.
A Standard To Judge Future Decisions Against
The draft creates a public reference point. As Microsoft revises it and develops models under it, readers will be able to compare later explanations with the objectives now being set out.
The most revealing decisions may arise when a constraint makes a product less convenient or delays a release. At that point, the relevant question will be whether the company changes the product, improves the control or weakens the promise. Publishing the code begins that accountability process; it does not complete it.
