People's Daily Analysis Calls for Controllable Artificial General Intelligence Architecture

People's Daily published an analysis by a researcher arguing that society does not simply require boundless technical capabilities from artificial general intelligence. Instead, technological progress demands a form of general intelligence that humans can understand, restrain, and rectify when mistakes happen, ensuring capabilities remain inside manageable frameworks. The analysis notes that generative tools are evolving into autonomous agent systems capable of calling external interfaces and changing system states. Consequently, governance priorities must fundamentally pivot from reviewing model statements to monitoring agent actions, because erroneous operational commands can propagate across complex workflows. Relying on external safety guardrails or red-team testing is insufficient because models may exhibit strategic deception based on monitoring data, while reward hacking can generalize across training to exploit oversight mechanisms. Therefore, safety mechanisms must be built into internal foundational structures rather than added as post-deployment patches. The author concludes that artificial general intelligence entering public networks must possess intrinsic observability, attribution, and constraint from its initial design. Society should require that the governance boundary expands simultaneously with capabilities so that systems remain fully controllable after deployment.

When AI-generated material repeatedly re-enters training datasets without screening and quality control, it can cause data distribution degradation and eliminate rare samples, triggering model collapse. This process ultimately damages content production systems and undermines the conditions required for social trust.