Nadella Aims to Contain AI by Assuming Compromised Models

The CEO of Microsoft calls for treating every AI model as potentially compromised and to contain it from the outset. He advocates for constant human oversight and standardized containment technologies, in addition to transparency requirements already shared within the industry.

Default Containment and Human Kill Switch Required

Satya Nadella asserts that an AI model must be contained from the start, operating under the assumption that it is compromised. He compares this containment to an emergency brake, believing that an authorized person should always be able to pause or stop a model while it is in operation. According to him, more advanced models will require more sophisticated containment technologies, which should be standardized. He emphasizes the necessity of assuming that all AI models are compromised.

Transparency and Security Practices Demanded

Satya Nadella advocates for a more open framework that allows for the management of AI models, monitoring them, and generating human-readable and tamper-proof audit trails. He also calls for immediate incident notifications, external checks, the use of verifiable data, and the isolation of models. Some of these recommendations align with positions already expressed in the industry.

Rejection of Black Boxes and Context of the Statement

Satya Nadella believes it is no longer acceptable to view AI as a set of black boxes from which advice and actions are simply accepted or rejected. He articulated these positions in a lengthy message posted on X as the CEO of Microsoft. AI is also referred to as superintelligence.