Do I need to file if I use an overseas LLM API?
Short answer
Calling an overseas LLM API does not trigger LLM filing in China, because you are not training or substantively modifying the model itself. However, depending on how the API is used, other obligations may still apply: algorithm filing, data export assessment, and content safety management.
Boundary conditions
LLM filing — generally not triggered. Filing applies to entities that train or substantively modify a large model. Calling an API from a provider such as OpenAI, Anthropic, or Google falls outside this scope.
Algorithm filing — may be triggered. If your product uses the model output in a recommendation feed, ranked list, search result, or personalised scheduling system, algorithm filing obligations apply regardless of whether the model is domestic or overseas.
Data export — may be triggered. If user prompts, uploaded files, or personal information are sent to the overseas API, this constitutes data export. Whether a security assessment, standard contract, or certification path applies depends on data volume and sensitivity.
Content safety — always required. Even if the model provider has its own safety filters, you are responsible for the content shown to Chinese users. You need your own content review layer, not just the model provider’s built-in safety.
Generative AI service registration — may be required. If you provide generative AI services to the Chinese public through an overseas model API, you may still need to complete the generative AI service registration, depending on the service scope and user base.
Next steps
Map the data flow first. Draw out exactly what data goes to the overseas API, what comes back, and who sees the output. This determines which obligations apply.
Check the data export path. If personal information is involved, determine whether you fall under security assessment, standard contract, or certification. The standard contract path is the most common for API-based products.
Add a domestic content safety layer. Do not rely solely on the overseas model’s safety filters. Implement your own input filtering and output review for content served to Chinese users.