Provisions on the Administration of Algorithmic Recommendations
互联网信息服务算法推荐管理规定, Order No. 9 of four departments. Effective 1 March 2022. China's core platform-algorithm rule, read in full from the official text.
In one line
Whoever pushes information to users by algorithm must take responsibility for the algorithm itself: publish how it works, let users switch it off, file it with the regulator — and never use it to manipulate opinion, addict children, or price-discriminate.
Who and what it governs
- Applies to providing internet information services in China by applying recommendation technology (Article 2).
- Those technologies are defined broadly: generation and synthesis, personalised push, ranking and selection, search and filtering, scheduling and decision (Article 2).
Key obligations
| Article | Obligation |
|---|---|
| 7 | Algorithm safety responsibility: audit mechanisms, ethics review, user registration, publication review, data and personal-information protection, anti-fraud, security assessment, incident response; publish service rules; staff the service properly. |
| 8 | Regularly review and validate the algorithm; must not build models that induce addiction or over-consumption. |
| 9 | Maintain a feature library of illegal and harmful information. Content generated or synthesised by an algorithm must be prominently labelled before it may be transmitted — the hook into the labelling measures. Illegal information must be stopped and reported. |
| 10 | Manage user models and tags; illegal or harmful keywords must not be recorded as interests or tags. |
| 11 | Ecological management of prominent positions: front page, trending, featured, rankings, pop-ups. |
| 14 | No fake accounts, account trading, account manipulation, fake likes/comments/shares, no information blocking, over-recommendation, ranking manipulation, or control of trending or featured lists. |
| 15 | No unreasonable restrictions on other service providers; no monopoly or unfair competition. |
| 16 | Tell users prominently that recommendation is used, and publish the basic principles, purpose and main operating mechanism. |
| 17 | Give users a non-personalised option or an easy way to switch recommendation off, and a way to select or delete personal-feature tags. Switching off must take effect immediately. |
| 18 | Minors: age-appropriate modes; no content that could prompt unsafe imitation or bad habits; no algorithmic inducement into addiction. |
| 19 | Elderly users: age-friendly services, plus monitoring and handling of telecom and online fraud. |
| 20 | Workers under scheduling algorithms: order allocation, pay structure, working hours, rewards and penalties. |
| 21 | Consumers: no unreasonable differential treatment in price or other terms based on preference or trading history. |
| 22 | Accessible complaint and reporting channels with clear process and response deadlines. |
Filing and supervision
- Article 24 — a provider with public-opinion attributes or social-mobilisation capacity must file through the algorithm filing system within 10 working days of starting service, reporting its name, service form, application field, algorithm type, self-assessment report and proposed public disclosure. Changes: within 10 working days. Termination: cancel the filing within 20 working days.
- Article 25 — the regulator files and publishes a filing number within 30 working days.
- Article 26 — the filing number must be displayed prominently with a link to the published information.
- Article 27 — such providers must also undergo a security assessment.
- Article 23 — tiered and classified management based on public-opinion attributes, content category, user scale, data importance and the degree of intervention in user behaviour.
Penalties
- Article 31 — for breaches of Articles 7, 8, 9(1), 10, 14, 16, 17, 22, 24, 26: warning, public criticism, order to correct within a deadline; if refused or serious — suspend information updates and fine RMB 10,000 to 100,000.
- Article 32 — for breaches of Articles 6, 9(2), 11, 13, 15, 18, 19, 20, 21, 27, 28(2): handled under the relevant laws and administrative regulations.
- Article 33 — obtaining a filing by concealment or false materials: revocation, warning, public criticism; serious cases: suspend information updates and fine RMB 10,000 to 100,000.
Relationship to other instruments
- Built on the Cybersecurity Law, Data Security Law, Personal Information Protection Law and the Measures for the Administration of Internet Information Services (Article 1).
- Article 9 connects directly to the AI content labelling regime.
Unverified / TODO
- Note the implementing guidance for the algorithm filing system
- Cross-reference the content-ecology rules invoked in Article 9