The Cyberspace Administration of China released draft Interim Measures on the Administration of Human-like Interactive Artificial Intelligence Services for consultation on 27 December 2025. The measures target companion and persona style systems, the category that sits closest to users emotionally and furthest from the enterprise compliance perimeter. Surveys of the Chinese framework published in August 2026 record the measures as having taken effect on 15 July 2026.
The disclosure requirement is blunt. Services must show conspicuous alerts telling users they are interacting with AI and not a natural person. There is no design exemption for immersion.
The crisis handling requirement goes further than anything in force elsewhere. Providers must run a tiered intervention framework: automated supportive responses when risk is detected, and mandatory human operator takeover when a user expresses self harm or suicidal intent. Escalation to a person is not a best practice in this regime. It is a licensing condition.
Minors get a dedicated regime. Guardian consent is required, a minors mode with usage limits must exist, parental monitoring must be available, and the service must switch into that mode automatically once a minor is identified. Elderly users get their own protections: emergency contact setup is required, and services are prohibited from simulating family members. That last clause is specific enough to name the fraud pattern it exists to stop.
Operationally, providers must maintain management systems spanning algorithms, content, cybersecurity, data security and emergency response, embed safety measures into their data pipelines against state defined values and cultural standards, conduct security assessments with reports to provincial cyberspace authorities under specified conditions, and comply with the existing algorithmic recommendation filing process. Enforcement runs from warnings and reprimands to rectification orders and suspension of service.
The measures do not sit alone. The Cyberspace Administration ran a four month AI application rectification campaign during 2026, publishing first stage results on 6 July and second stage results on 2 September, with action taken against products, accounts and platforms over filing failures, safety review failures, data poisoning and content labelling violations. Enforcement is being reported as a running campaign rather than as isolated cases.
Above all of it, Beijing is drafting a single comprehensive AI law intended to consolidate treatment of data, algorithms, computing power, cybersecurity and intellectual property. Chinese technical standards work has moved in parallel, with guiding documents issued in 2026 covering AI agent interconnection architecture and agent technical requirements.
The pattern is worth reading precisely because it inverts the Western sequence. The European Union wrote a horizontal statute and is building enforcement machinery underneath it. China wrote narrow, behaviour specific rules for deployed products first, is enforcing them in campaigns, and is assembling the general statute afterward from what the sectoral rules taught it.
For any vendor shipping a conversational product into the Chinese market, the compliance surface is not abstract. It is a disclosure banner, a staffed escalation desk, an age gate with a mode switch, and a filing with a provincial regulator.