Microsoft’s Nadella calls for an emergency brake on AI models

Controls outside the model

In his post, Nadella said it was time to stop and assess the architecture of trust in AI. He used “superintelligence,” a term preferred by the Trump administration, to describe AI.

His proposal has three parts:

Nadella spoke amid reports from leading companies of more frequent cases in which they appeared to lose control of their models. Anthropic CEO Dario Amodei had earlier published a plan for more cautious AI development.

The missing test

I think the most consequential part is not the shutdown switch but the demand for a legible record of what the model did. A control is only useful if people can see when it should be used; Nadella’s proposal puts that visibility alongside the ability to intervene.

What the post does not specify is how to decide which actions count as significant, or who gets authority to stop a model. Those details determine whether an emergency brake is a practical safeguard or a reassuring metaphor. The architecture Nadella describes makes oversight possible in principle; its value will depend on whether the people responsible can understand the record and act in time.