What is the difference between algorithm filing and LLM filing?
Short answer
Algorithm filing governs how content is selected and pushed to users — recommendation, ranking, search, and scheduling mechanisms. LLM filing governs whether the model and its training data are safe — whether the company developed the model in-house or made substantive modifications to it. Different legal basis, different triggers. They are not an either/or choice.
Boundary conditions
Both obligations can apply at the same time. When a product combines generative AI with content feeds, leaderboards, or search ranking, both filing obligations are triggered. Doing one does not cover the other; gaps will be flagged during inspections.
Calling a filed model without modifying its capabilities triggers registration, not filing. You submit entity and product information, the model’s filing number and authorisation, and your data processing and content safety measures. If the product also has algorithmic recommendation features, algorithm filing is still required.
Internal use only, no recommendation or distribution, no public-facing generation — generally does not trigger filing. But once you add a customer-facing website, chatbot, or mini-program, the audience becomes public and the assessment changes.
Face generation, voice cloning, and digital avatar driving are a third set of obligations. Deep synthesis filing applies separately; if the generation uses an in-house or substantively modified model, LLM filing is also required.
The starting point for assessment is the model source, not the business form. Self-training, continued pre-training, large-scale instruction fine-tuning, and changing the base model all change the model’s capability boundaries. Prompt engineering, retrieval-augmented generation, and UI wrapping usually fall on the registration side. Local interpretations of “substantive modification” vary, and the same technical approach can receive different conclusions in different regions.
Next steps
Confirm the classification before preparing materials. Before drafting submission materials, consult the local cyberspace administration about model source, training method, and data composition to confirm which path applies.
Reverse-check materials against common rejection reasons. Four items are the most frequent causes of returns: training data source and authorisation documentation, whether the safety assessment report matches the product’s actual functions, whether complaint channels are genuinely operational, and whether generated content implements both explicit and implicit labelling.