China MIIT Launches National AI Ethics Review Pilot — Provincial Committees to Address Algorithmic Discrimination and Emotional Dependence Risks

China's Ministry of Industry and Information Technology (MIIT) issued a formal notice on May 10, 2026 announcing the launch of a national pilot program for AI ethics review and services, beginning in provincial regions hosting national AI industrial innovation zones. The program establishes three primary mandates: creating provincial-level AI ethics review rules and standards; setting up formally constituted AI ethics committees with multi-stakeholder membership (government officials, industry representatives, academics, and civil society); and converting practical review experience into national technical standards for broader adoption. The pilot specifically targets two AI risk categories identified as regulatory priorities: algorithmic discrimination (including biased outcomes in hiring, lending, and public service allocation) and emotional dependence (risks from AI companion applications and emotionally manipulative recommendation systems). The program is the latest governance action building on China's August 2023 Generative AI Interim Measures — the world's first national framework specifically regulating generative AI. The ethics pilot comes four days before the Trump-Xi Beijing summit (May 14–15), where formal US-China AI dialogue is under consideration. China's move to institutionalize AI ethics review infrastructure simultaneously advances governance goals and strategic positioning — demonstrating that Beijing is developing its own AI regulatory standards at a national scale. The pilot reflects the 15th Five-Year Plan's 'AI safety governance' priority alongside better-publicized AI investment directives. MIIT has maintained that China's AI governance approach — mandating algorithmic transparency, content labeling, security assessments, and user notification — is substantively more comprehensive than Western voluntary commitments, even while embedding political constraints on model content.