China's Cyberspace Administration opened a unified filing portal at beian.cac.gov.cn on November 9, consolidating generative AI, algorithm, deep synthesis and blockchain registrations into a single system. The previous algorithm filing system and blockchain filing system were retired the same day.
For anyone shipping AI products to Chinese users, the URL change is the least important part. The substantive shift is that the categories now share one intake, which means the material requirements will converge and the ambiguity about "which system applies to me" shrinks.
Three different obligations, often confused
Filing (备案). Providers that offer generative AI services to the domestic public with public-opinion attributes or social mobilization capacity must file under the Interim Measures for the Management of Generative AI Services. This is for organizations training foundation models and serving them directly.
Registration (登记). Applications that call an already-filed model through an API or similar route and serve the domestic public register rather than file. Provincial authorities publish batches: Sichuan added five registered services on November 2, Hebei announced four on November 11.
Labeling (标识). Under the Measures for Labeling AI-Generated Synthetic Content, live generative AI services must add labels to generated content and display the model name and registration number prominently, or on the product detail page.
These map to different responsible parties: the model provider, the application integrator, and the content publisher. Most teams that get stuck have not decided which layer they are on.
A pre-launch checklist
Confirm your role. Self-trained model, API integrator, or content distributor. The obligations differ; do not copy another company's filing pack.
Check the disclosure surface. Product detail page, about page, generation result page: is the model name and registration number there? This is the most commonly missed item and the easiest to audit.
Test the label chain end to end. Visible labels on text, image, audio and video; metadata labels written; labels surviving export, share and re-edit. Test it, do not assume it.
Prepare the core documents. Algorithm mechanism description, safety assessment, corpus provenance and legality, user agreement, complaint channel. These are standard and slow to assemble late.
Keep logs. User input, generated output, moderation decisions, enforcement actions. Retention period follows risk. When something goes wrong, logs decide whether this is a management defect or negligence.
Four traps
Treating registration as a lighter version of filing and submitting thin material. Registration is reviewed too.
Assuming only large platforms are in scope. Universities are tightening as well: one institution required internal agents to file before launch, including research-group chat tools.
Ignoring change and deregistration. Model upgrades, service shutdowns, changes of operating entity all need updates, or you leave a violation on record.
Turning the label into a disclaimer buried in a collapsed menu. The label exists so users know content is machine-generated. Placed at the bottom of a settings drawer, it effectively does not exist.
Compliance is not a one-time event; it tracks product iteration. A new portal is a good excuse to re-audit the four things that actually get checked: filing status, disclosure, labeling and logs.
China's New AI Filing Portal Is Live. Here Is What Teams Must Submit
← Previous: Embodied AI's Funding Week: UniX AI's $70M and DeepSeek's Two Bets
Next: →Jensen Huang's Last Argument: Compete With China, Don't Withdraw →