Recommendation algorithms and the filing system
Lesson 2 of 5 in China: The World’s Most Developed Binding AI Rulebook.
The stack starts with the Provisions on the Administration of Algorithmic Recommendation of Internet Information Services, effective 1 March 2022 — the world’s first binding, economy-wide rules on recommender systems. They were born of two very Chinese anxieties: viral content the state cannot steer, and platform business practices that had become a public scandal.
The scandal had a name: “big data killing familiarity” (dàshùjù shāshú) — platforms charging loyal customers more than new ones because the algorithm knew they would pay. The provisions banned it outright: no unreasonable differential pricing based on user preferences or transaction habits. They also reached the gig economy, requiring platforms that schedule workers by algorithm — think food-delivery riders racing timers set by a model — to build in rest, reasonable pay, and transparency about how orders are assigned.
User rights: know, refuse, and delete
Users must be told when algorithmic recommendation is used and offered a way to switch off personalisation entirely. They may also delete the tags a platform’s profile has attached to them. In practice every major Chinese app now ships a 关闭个性化推荐 (turn off personalised recommendation) toggle — a visible, testable compliance artifact.
Protected groups: minors, the elderly, workers
Providers may not use algorithms to induce minors into addiction or unsafe behaviour; services for the elderly must account for their needs and guard against telecom fraud; gig workers scheduled by algorithm get statutory protection for pay and rest. This is the consumer-protection face of the provisions — obligations with no EU AI Act equivalent until you reach into labour law.
Content duties: steering the feed
Providers must promote positive energy content and must not use algorithms to spread information the law prohibits, manipulate rankings, fabricate accounts, or hijack traffic. Here the information-control DNA is explicit: the state regulates not just what may be said but what the algorithm may amplify.
Synthetic content: the first labeling seed
Already in 2022, the provisions required that algorithmically generated or synthesised information be conspicuously labeled — the seed that grew into the deep-synthesis rules (2023) and the full labeling regime (2025).
The provisions also created the mechanism that now threads through every Chinese AI rule: the algorithm filing system (算法备案). Providers whose algorithms have “public opinion attributes or social mobilization capability” must file with the CAC within ten working days of launch: the provider’s name, algorithm type, a self-assessment report, and the content to be displayed. The CAC publishes the registry — thousands of filings are now public, giving researchers a rare window into China’s algorithm landscape.
Do not mistake filing for a rubber stamp. A filing can be rejected or revoked, and launching a covered service without one is a sanctionable violation — which, combined with the security-assessment requirement you will meet in the generative-AI lesson, makes market entry for consumer AI in China a permissioned event. Western analysts describe the combined effect as licensing in all but name.
Must you file your algorithm with the CAC?
Interactive decision tree — outcomes:
- Largely outside the service rules
Internal enterprise tools and pure R&D are carved out of the public-facing service regulations (the GenAI Measures say so expressly). PIPL, the Data Security Law, and ethics-review duties may still apply — but no algorithm filing is triggered.
- Provisions apply; filing does not
You still owe the substantive duties — opt-out toggles, no price discrimination, labeling of synthetic output — but the filing obligation attaches only to algorithms with public opinion attributes or social mobilization capability. Assess honestly: the CAC reads this trigger broadly.
- File within 10 working days — and expect scrutiny
File the algorithm with the CAC registry (provider, algorithm type, self-assessment report). If the service is also generative AI facing the public, a security assessment comes first. Treat launch as permissioned, not notified.
- You are exposed
Operating a covered algorithm without filing is a sanctionable violation: expect rectification orders, suspension of new-user registration, app-store removal, and fines. CAC enforcement campaigns actively sweep for unfiled services.
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.