Algorithmic recommendation and filing

Must you let users switch algorithmic recommendation off, and how do you tell them?

What you may need to do

1. Put a convenient algorithmic recommendation off switch in the settings page
2. Stop personalised recommendation immediately after the switch is off, and serve non-personalised content
3. Provide an entry point to view, select and delete the user tags based on personal characteristics
4. Inform users in a prominent way that the service is provided, and disclose the basic principles and operating mechanism

Clause basis

One-line answer

Users cannot be left on “accepted by default” for algorithmic recommendation. A business must inform them in a prominent way and offer either a non-personalised option or a convenient off switch; once the user turns it off, the relevant service must stop immediately, and the user must also be able to select or delete the user tags based on personal characteristics.

The basis

Article 16 governs notice and disclosure: providers of algorithmic recommendation services shall inform users, in a prominent way, that they provide an algorithmic recommendation service, and shall disclose, by appropriate means, the basic principles, the purpose and intent and the main operating mechanism of the service.

Article 17 governs the right to choose: paragraph 1 requires an option that is not directed at the user's personal characteristics, or a convenient option to turn the algorithmic recommendation service off; where the user chooses to turn it off, the relevant service must stop immediately. Paragraph 2 requires the function of selecting or deleting the user tags based on personal characteristics that are used for the service. Paragraph 3 adds that where applying the algorithm has a major impact on users' rights and interests, the provider shall explain it in accordance with the law and bear the corresponding responsibility.

Alongside these sits Article 22, the appeal and complaint channel: providers of algorithmic recommendation services shall set up convenient and effective entry points for user appeals and public complaints and reports, make the handling process and the reply time limit clear, and accept, handle and reply promptly.

What you need to do

• Build a personalised recommendation switch in the settings page of the app or website, in a place reachable within a few steps.
• After the switch is off, serve the non-personalised channel (sorted by time or popularity, for example) rather than swapping in another personalisation model.
• Provide a tag management page: show the personal-characteristic tags used for recommendation, and support selecting and deleting them.
• Write the basic principles, purpose and intent and main operating mechanism of the algorithmic recommendation service as an explanation aimed at ordinary users, and publish it.
• State the handling process and the reply time limit for the user appeals entry point, and keep a record of what is received.

Common pitfalls

Making the switch hard to find, or showing a pop-up that keeps urging the user to turn the service back on after they have turned it off, conflicts with the requirements of “convenient” and “stop immediately”. Building the switch but not tag management fails an obligation that Article 17, paragraph 2 imposes separately. Turning the notice into a one-off pop-up at registration leaves users unable to find the operating mechanism explanation later. Deleting a tag while the recommendation still builds a profile from the same characteristic is no deletion at all. An appeals entry point that is only an email address with nobody handling it makes the reply time limit meaningless.

Common follow-up questions

Whether leaving the recommendation area empty after a user turns recommendation off is compliant is not directly prescribed; it can be understood as a need to keep the service usable. For the specific approach, we suggest talking to the local cyberspace administration.

Common follow-up questions

After a user turns algorithmic recommendation off, may we still sort by popularity?

Article 17, paragraph 1 requires an option that is not directed at the user's personal characteristics. Organising content by popularity or by time — methods that are not directed at personal characteristics — fits that direction.

How much granularity must the user tags be shown at?

The article requires the function of selecting or deleting them and does not prescribe the granularity of display. We suggest a level the user can recognise and act on; when in doubt, follow the practice of the local cyberspace administration office.

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