Long before "generative AI" entered every regulator's vocabulary, Chinese users were already being sorted, ranked, and priced by invisible code. The rule that first named that code as a thing to be governed was the Provisions on the Management of Algorithm-Recommendation Services for Internet Information Services (《互联网信息服务算法推荐管理规定》).
It took effect on March 1, 2022. Most people outside China's policy circle have never heard of it — yet it quietly reshaped what every major app must show, and what it must let you turn off.
A rule aimed at the algorithm itself
The provisions (Order No. 9) were issued jointly by four agencies: the Cyberspace Administration of China (CAC, 国家网信办), the Ministry of Industry and Information Technology (MIIT, 工信部), the Ministry of Public Security (MPS, 公安部), and the State Administration for Market Regulation (SAMR, 市场监管总局).
They were adopted on November 16, 2021 and promulgated on December 31, 2021. Chinese commentary at the time called it the country's first departmental rule with the algorithm as its specific object of regulation.
Which algorithms it covers
The text defines "algorithm-recommendation technology" to include five families:
- generative-synthesis (生成合成类)
- personalized-push (个性化推送类)
- ranking-and-feature (排序精选类)
- retrieval-and-filter (检索过滤类)
- scheduling-and-decision (调度决策类)
That list captures almost the entire machinery of a modern platform: what you see in a feed, what you are offered to buy, how a ride or a delivery is dispatched, and how a search result is ordered.
The user-rights core
Three requirements are the ones ordinary users actually feel:
- Turn it off. Providers must offer a clear option to disable personalized recommendation, and must stop personalization immediately when chosen.
- See and delete your tags. Users can view and delete the personal-feature tags used for recommendation.
- No "big-data price discrimination" (大数据杀熟). Pricing by algorithm based on a user's profile or habits is prohibited as an unreasonable differential-treatment practice.
The law also protects specific groups: minors (no addictive or harmful pushes), elderly users (accessible, age-appropriate service), workers (fair scheduling and pay for delivery and similar labor), and consumers (fair-transaction rights).
The filing system
For algorithms with "opinion-attributed or public-mobilization capability" (舆论属性或者社会动员能力), providers must file (算法备案) with the CAC within ten working days of launching, and display the filing number on the service.
This algorithm filing (算法备案) system was, and remains, the practical enforcement hook: it makes the existence of a ranking or push algorithm a registered fact rather than a black box.
Penalties
Violations can draw warnings, rectification orders, and fines in the range of RMB 10,000–100,000 (≈ US$1,400–14,000), with suspension of information updates for serious or repeated cases, and criminal referral where applicable.
Why it still matters in 2026
Two reasons. First, the same agencies later extended the logic to AI-generated content (the 2025 labeling measures), so the 2022 algorithm rule is the older root of China's "register and disclose" approach to automated systems. Second, recommendation algorithms did not go away when chatbots arrived — they still govern feeds, rankings, and prices, and this rule is still the binding one for them.
What platforms actually had to change
The obligations read like a checklist for any feed, store, or dispatch system:
- Maintain algorithm-security responsibility: mechanisms for review, ethics review, and user registration.
- Label undisclosed synthetic information before transmitting it (Article 9).
- Do not manipulate hot-search lists, rankings, or trending boards to steer public opinion (Article 14).
- File, within ten working days, and display the filing number.
In practice this touched short-video apps, e-commerce platforms, news aggregators, and food-delivery and ride-hailing dispatch systems — anywhere a ranking or push algorithm decides what a user sees or earns.
The "big-data price discrimination" clause
One provision got the most public attention: a direct prohibition on using a user's preferences or transaction habits to apply unreasonable differential pricing — the practice Chinese users call "big-data 杀熟" (big-data rip-off). State-media explainers tied this to real disputes, including a widely reported hotel-booking case where a platform member was charged nearly double the hotel's own rate. The rule did not invent the lawsuit remedy, but it gave regulators an explicit statutory hook.
How it connects to later AI rules
The same four agencies that issued this 2022 rule later issued the 2025 AI-generated-content labeling measures. The through-line is consistent: China's approach has been to register and disclose automated systems first (algorithms in 2022, synthetic content in 2025), rather than to ban by model capability. For a foreign reader, that is the key contrast with the EU AI Act's risk-tiering of the models themselves.
Honest limitations
This article is built from the official CAC text and the Chinese government gazette, plus contemporaneous state-media explainers. I did not independently audit enforcement statistics (e.g., how many filings exist or how often penalties were imposed). The "first of its kind" characterization is the CAC's own expert commentary, not an independent global comparison. I did not assess how the rule interacts with subsequent AI-labeling or generative-AI measures beyond noting the lineage.
What readers can do now
- If you build or ship software in China, map your personalization, ranking, and dispatch features to the five algorithm types and check whether any carry opinion/mobilization attributes requiring filing.
- As a user, exercise the off-switch: locate the "personalized recommendation" toggle in major apps and test the difference in your feed.
- When comparing China's AI governance to the EU AI Act, note the different starting point — China's first binding AI-adjacent rule targeted recommendation algorithms in 2022, not foundation models.