AI Safety Governance Framework 3.0 (人工智能安全治理框架3.0)

Key Findings

• — The derivative-risk taxonomy names risks of emotional dependence from anthropomorphic AI services, with adolescents and older adults singled out, and a separate social-alienation risk from virtual intimate relationships displacing real-world interaction
• — Technical countermeasures for anthropomorphic services include overreliance-risk warnings and emotional-boundary guidance, with more prominent alerts such as pop-up warnings when users show signs of addiction or excessive usage duration
• — Guideline 5.3.17: AI services for minors, older adults and special groups must build anti-addiction and anti-over-dependence mechanisms into product design and service models
• — Guideline 5.3.20: providers of anthropomorphic interaction services should improve AI-identity reminders, usage-duration reminders and crisis-intervention mechanisms, and prevent excessive pandering to users, induced emotional dependence and emotional manipulation, protecting minors' physical and mental health
• — Embodied-AI companionship and care-assistance risks are listed separately: identity misperception, bodily-boundary violations, emotional dependence and manipulation, and unclear liability
• — Three appendices grade risks by scenario, intelligence level and deployment scale, set out an agentic-AI risk management framework, and list fundamental principles for trustworthy AI

Methodology Notes

Non-binding policy framework, not an empirical study; Chinese text followed by an English translation in one PDF (about 130 pages). Drafted by TC260 with the Chinese Academy of Cyberspace Studies, the CAC data and technology support centre, research institutes and industry firms; supersedes Framework 1.0 (2024) and 2.0 (2025). Release date from the CAC announcement of 14 September 2026; the PDF cover is dated September 2026. Curator fetched the bilingual PDF from cac.gov.cn (3.3 MB) on 2026-09-17 and read clauses 3.3, 5.3.17 and 5.3.20 in both languages.

Sources