ME AI News: Microsoft CEO Satya Nadella published a lengthy post saying that as superintelligence systems develop, the industry needs to rethink its AI trust architecture. It cannot simply rely on assurances from model providers, nor can it outsource responsibility for AI behavior. Nadella said that when traditional software systems were deployed across industries over the past several decades, developers could usually trace behavior to specific code paths. Today, however, the capabilities of frontier AI models have surpassed those of traditional software systems, yet their outputs cannot be clearly attributed to specific training data or model-weight configurations. Companies are connecting autonomous AI systems to sensitive data and granting them permission to carry out critical tasks, so they need to build systems that are observable, testable, and controllable, separating the “supply of intelligence” from “access control.” Nadella proposed governing superintelligence through an engineering approach: constrain non-deterministic models with deterministic system design, human oversight, and operational procedures, and manage AI systems as potential “insider risks”—including by limiting permissions, logging behavior, and establishing isolation boundaries. He also said superintelligence systems should follow principles such as multi-model collaboration, end-to-end observability, continuous validation, independent control, independent auditing, risk isolation, and incident disclosure. He emphasized that transparency into a model’s chain of thought (CoT) is a baseline requirement, but CoT transparency alone is not enough, since model outputs may still lack reliable explainability. Nadella concluded that the most trustworthy superintelligence systems of the future will not be “the most trustworthy models,” but “systems that can still operate safely even when the models are not fully trusted.” (Source: ME)