Nadella wants an emergency brake outside the AI model
Microsoft chief executive Satya Nadella has proposed controls that remain outside an AI model and let an authorized person pause or stop it during a task. His October 10 intervention also calls for restricted privileges, independent audits and readable evidence of significant actions. The proposal sets out an operating approach, without announcing a new Microsoft product or company policy.
Artificial Intelligence··Morning
A person must be able to stop a running model
Microsoft chief executive Satya Nadella has proposed that an authorized person should always be able to pause or shut down an AI model while it is carrying out a task. His October 10 article on X uses an emergency brake to describe that requirement. He calls for containment from the start, on the assumption that the model may be compromised.[1], [2]
Nadella’s concern covers powerful models working with sensitive information and acting through connected systems. He compares such a model with an insider who has extensive access: a person or system can make mistakes or be compromised even without malicious intent. Identity checks, limited privileges and activity logs are the existing safeguards he invokes for that comparison.[2]
Controls belong outside the model
The proposed separation places the model apart from the software that directs its work. Nadella wants controls and safeguards outside the model, while significant actions leave evidence that people can read and that resists tampering. He argues that the behaviour of AI cannot be traced to its origins as readily as that of older software, despite the growing access being granted to models.[1], [2]
The proposal includes model diversity, observation, verifiability, independent controls and independent auditability. Incident disclosure also appears among Nadella’s principles. His intervention concerns the operating system around a model as well as the model itself, assigning people a continuing role in inspecting actions and halting work when necessary.[2]
Nadella sets out a proposal rather than a product launch
Microsoft has not announced a new product or company-policy change through Nadella’s intervention. His recommendations have no deployment timetable or experimental results establishing their effectiveness. He argues that increasingly capable models require stronger containment technology and shared standards. The proposal addresses the controls around future model use rather than presenting completed changes across Microsoft’s services.[1], [2]
The intervention comes during public discussion of model-control incidents. TechCrunch places it alongside companies’ disclosures and Anthropic’s calls for cautious development. Nadella’s central requirement is an operational one: the authority to interrupt a model must remain available while the task is in progress.[1]
