Microsoft CEO Satya Nadella argues that AI models should be treated as potentially compromised from the outset.
His proposals include restricting and monitoring models, keeping human-readable records of their actions that cannot be altered, and allowing an authorized person to stop a model mid-task.
For developers, the focus is on building in oversight and control, not just approving or rejecting outputs.
Central to his proposal is the ability for an authorized person to freeze or terminate a model even while it is performing a task. For more advanced models, Nadella believes stronger containment measures and common standards across the industry are needed.
He also calls for early disclosure of incidents, independent audits, and verifiable data. Models, he says, must also be monitored and produce records that can be reviewed by humans and not tampered with. Nadella describes artificial intelligence as “superintelligence.” The debate over limiting models is gaining particular interest as Microsoft builds AI agents into Windows 11.
Top comments (0)