Algorithmic recommendation and filing

Our app has a recommendation feed — does this algorithmic recommendation rule apply to us?

What you may need to do

1. Inventory every ordering, recommendation, filtering and dispatch logic in the product and tag each one
2. Decide whether each logic provides users with information, goods or service dispatch
3. Put the conclusion on whether information services are provided to the domestic public in writing and keep it on file
4. Consult the cyberspace administration authority for your locality about doubtful algorithmic scenarios before launch

Statutory basis

One-sentence conclusion

As long as algorithmic recommendation technology is applied to provide internet information services to users within the People’s Republic of China, the provisions apply, regardless of the size of the company, of whether a fee is charged and of whether personalisation is used.

Basis

Article 2 sets the scope very widely: applying algorithmic recommendation technology to provide internet information services within the People’s Republic of China (referred to as algorithmic recommendation services) is governed by these provisions, and only where laws or administrative regulations provide otherwise do those other provisions apply. The second paragraph of the same article defines the technology: algorithmic recommendation technology as referred to in the preceding paragraph means the use of algorithmic technologies of the generative-synthesis, personalised push, ranking and selection, retrieval and filtering, and dispatch and decision-making categories to provide information to users.

In other words, at least five categories of technology are named: home-page information feeds, short-video recommendation, search result ordering, leaderboards and selections, and dispatch and scheduling. Many teams assume that not doing personalised recommendation keeps them safe; but if any link in the product orders, filters, recommends or dispatches, it may fall into the retrieval and filtering or the dispatch and decision-making category.

The regulatory structure is also set out in Article 3: the national cyberspace administration department coordinates algorithmic recommendation service governance and the related supervision and administration nationwide; local cyberspace administration departments coordinate the work within their own administrative areas; and the relevant departments of the State Council and of local governments in telecommunications, public security and market regulation each handle supervision and administration within their own duties.

What you need to do

• Go through the ordering, recommendation, filtering and dispatch logic in your own product first, and tag each item.
• Decide whether each logic provides users with content, goods or service dispatch.
• Put the conclusion on whether the product provides information services to the domestic public in writing and keep it on file for inspection.
• For interfaces you cannot judge with confidence, consult the cyberspace administration authority for your locality before deciding the launch timetable.

Common pitfalls

Treating the absence of a recommendation algorithm in the product as an exemption, when default back-office ordering, trending lists and filters are algorithmic recommendation technology too. Mistaking a system used only by internal staff and not offered to the public for a covered case, and then in turn missing the mini-program, the web version and open interfaces. Some enterprises review only the main app and overlook the same algorithm being output to third-party channels through an interface — while the provisions are framed by the provision of a service, not by the client application.

Follow-up questions

This knowledge base only retrieves and reports the text of the provisions; it does not replace formal legal advice. Where the characterisation of a specific product is in question, put the product description in writing and consult the cyberspace administration authority for your locality.

Follow-up questions

Does an internal-only retrieval tool also have to comply?

The condition of application in the provisions is the use of algorithmic recommendation technology to provide internet information services. Where the system provides no service to the domestic public, whether it is covered has to be judged on its specific form; when in doubt, the position of the cyberspace administration authority for your locality governs.

Do popularity lists and default ordering count as algorithmic recommendation?

The retrieval and filtering, and ranking and selection categories named in the second paragraph cover ordering and filtering logic. Treat them as within range and document your reasoning.

This entry belongs to the China AI Compliance & Filing cluster (42 entries in total): View all · Answers home