China’s provisional rules for generative AI services lay out requirements for providers serving users in the country, including API providers. The framework covers what services may generate, how they should protect users, and what information regulators can examine.
Providers face content and licensing requirements
The rules require generative AI providers to adhere to core socialist values. Prohibited content includes pornography, terrorism, racism and material that threatens China’s national security.
Services whose algorithms can influence public opinion must be registered with the relevant authority. Providers should also obtain an administrative license in accordance with the law, although the document does not specify which providers are required to do so.
These obligations place content moderation and regulatory compliance at the center of operating a generative AI service in China. Providers must consider both the material their systems produce and the processes used to manage it.
User protections include bias and youth safeguards
The rules say algorithms must not discriminate based on ethnicity, gender, age, occupation or health. They also prohibit using algorithms for anti-competitive behavior.
For underage users, providers are encouraged to create an anti-addiction system similar to those used in video gaming. The document frames this as a user protection measure, while leaving the design of such a system to providers.
Together, the provisions address potential harms in how AI systems treat users and how younger people engage with them. They establish broad expectations without describing specific technical measures.
Providers must respond to illegal output
Service providers are responsible for identifying and stopping the generation of illegal content. After an incident, they are expected to correct the algorithms and report the matter to the relevant authority.
This approach makes the provider responsible for responding when a system produces illegal material, including content generated after a user prompt. The source article notes that prompts submitted to an image generator or chatbot could potentially lead to legal trouble for individuals.
Regulators can also request details about a generative AI model. The information listed includes its training data, size, type, tagging rules and algorithms, giving authorities visibility into how a service is built and operated.
China also calls for domestic AI capacity
The document links oversight with plans to expand the infrastructure and resources available for AI development. It calls for a public data training platform and the sharing of computing power.
Beijing has already proposed concrete rules for a state-backed, centralized platform that allocates public cloud resources based on customer needs. The national document also calls for “self-reliant innovation” in AI algorithms, frameworks, chips, software platforms and other infrastructure.
At the same time, it encourages “equal and mutually beneficial” international cooperation. The stated aim is to balance development and security, a formulation that speaks to the tension between encouraging AI growth and setting firm boundaries for providers.
The provisional rules therefore reach beyond content restrictions. They set expectations for licensing, user protections, incident response and regulatory access to model information, while calling for shared resources and stronger domestic capabilities.