One product can have several regulated roles
The 2023 interim measures apply to generative services offered to the public in China. Article 22 expressly includes programmable interfaces: describing a product as an API does not by itself take it outside the provider definition. Conversely, Article 2 excludes specified research and applications that do not offer generative services to the domestic public from these measures; it does not exempt them from all privacy, security or sector rules.
Treat this page as a relationship map anchored in the 2023 measures, not a claim that every AI use needs the same filing. A model supplier, application operator, content-sharing platform and app store can carry different duties within one commercial chain. Keep separate conclusions for each legal entity and function.
Read the rules as layers, not replacements
The dates below are commencement dates for binding instruments, except the agent opinions, whose date is publication. The 2025 labeling measures expressly refer back to the earlier service rules. The 2026 anthropomorphic measures add a function-specific layer, while the agent opinions are policy guidance. An operational filing announcement is evidence of administrative activity, not a new generally applicable AI statute.
| Instrument and date | Question it answers | Relationship |
|---|---|---|
| Algorithm rules · | How algorithms select, rank, generate or deliver information and protect users. | Underlying information-service layer; not confined to recommendation feeds. |
| Deep synthesis · | Synthetic media, technical support, content safety and labels. | A service/function layer retained alongside generative AI measures. |
| Generative AI · | Public generative services, training data, user inputs and governance. | Cross-refers to deep-synthesis labeling and algorithm filing. |
| Content labels · | Visible labels, file metadata, propagation and app distribution. | More specific production-to-distribution requirements, not a repeal of earlier rules. |
| Anthropomorphic interaction · | Sustained emotional interaction simulating human traits. | Special scope; not every assistant or chatbot. |
| Agent opinions · published | Decision permissions, supply chains and risk-based governance. | Guidance; no separate effective date stated. |
- Algorithm Recommendation Provisions, Articles 2, 7–9, 16–17 and 24
- Deep Synthesis Provisions, Articles 2, 7–20 and 25
- 2023 Generative AI Interim Measures, Articles 2, 7–17, 19 and 22–24
- 2025 AI Content Labeling Measures, Articles 2–12 and 14
- 2026 Anthropomorphic Interaction Measures, Articles 2 and 8–14
- Opinions on Regulated Application and Innovative Development of Intelligent Agents, items 4 and 6–11
Assign duties along the service chain
Do not use the upstream model supplier's statement as a substitute for checking the downstream product. A provider may remain responsible for user agreements, input protection, content handling and complaints even where a model is procured externally. The following evidence requests are editorial implementation suggestions, not a prescribed government checklist.
| Role | Rule locator | Practical evidence |
|---|---|---|
| Public generative-service provider | 2023 measures, Articles 7–15 and 22. | Training provenance, input-retention policy, complaint and content-response records. |
| Deep-synthesis provider or technical supporter | Deep synthesis, Articles 14–19. | Training-data safeguards, biometric-editing consent prompts, labeling and applicable assessment/filing analysis. |
| Content dissemination service | Labeling measures, Article 6. | Metadata detection, user declaration, suspected-content labels and dissemination metadata. |
| App distribution platform | Labeling measures, Article 7; deep synthesis, Article 13. | Service declaration and review of relevant labeling, assessment and filing materials. |
Test labels through export and redistribution
Under Articles 4–5 of the labeling measures, distinguish a visible notice from implicit file-metadata information. A watermark is not a substitute for checking both applicable duties. Test the actual download, copy and export paths, not only a screenshot of the generator. Article 4 requires qualifying exported files to contain the required visible labels.
Article 6 distinguishes confirmed metadata, a user declaration without detected implicit metadata, and suspected generation traces. The platform's notice should reflect that evidence rather than mislabel every unmarked file as authentic. Article 7 adds app-distribution review; Article 12 connects labeling materials to applicable filing and assessment procedures.
Article 9 has a specific route for a user-requested output without a visible label: clarify the user's labeling obligations and responsibility through an agreement and retain relevant recipient logs for at least six months. It is not a general permission to strip implicit metadata or conceal generated content. Article 10 separately governs user declarations and prohibits malicious removal, alteration, forgery or concealment.
Separate filing, security assessment and operational notices
Article 17 of the generative AI measures links security assessment and algorithm filing to services with public-opinion attributes or social-mobilization capability. Deep-synthesis Articles 19–20 contain related provider and technical-support filing provisions and a new-product/function assessment rule. Separately, deep-synthesis Article 15 addresses assessment of specified biometric-editing and other sensitive synthesis tools. A single generic 'AI filing required' label hides these different triggers.
Record the exact service, entity, functionality and version covered by a public filing entry. An entry is not government approval of every output, a blanket privacy clearance or proof that every downstream deployment is covered. Changes to features, audiences or delivery methods require renewed applicability analysis rather than copying a vendor's filing number.
An assistant is not automatically an emotional companion
Article 2 of the 2026 measures targets public services in China simulating human personality, thinking and communication for sustained emotional interaction, including emotional care, companionship and support. It excludes customer service, knowledge Q&A, work assistance, education and research without sustained emotional interaction. The test is the actual interaction design, not whether a marketing page uses the word 'assistant'.
For covered services, Articles 8–14 address manipulation, overdependence, lifecycle safety, training, agreements, extreme emotional risks and minors. Article 14 prohibits providing minors with virtual intimate-relationship services and requires guardian consent for other covered services to children under 14. Do not interpret this as a ban on every educational chatbot or as permission to disregard broader child-protection rules.
Agents add an execution-permission question
The May 2026 agent opinions distinguish decisions reserved to the user, delegated decisions and autonomous decisions. Item 6 preserves the user's knowledge and final decision and prohibits exceeding authorization. Items 8–10 address permissions, attacks, model/API/tool supply chains and human intervention. These are policy directions to map against applicable law and sector requirements, not a new universal AI licence.
Item 4 explores an agent-registration platform; it does not announce that every agent must already register there. Item 11 distinguishes sensitive sectors from lower-risk scenarios and refers to applicable regulation. For an implementation review, test permissions at the actual action boundary: sending, purchasing, deleting, exporting or changing records should not inherit unlimited authority merely because a conversation began.
| Original control suggestion | Evidence | Failure to test |
|---|---|---|
| Limit tool permissions to the task. | Permission matrix and action log. | A document prompt attempts to trigger an unauthorized external send. |
| Confirm consequential actions. | User-visible action details and confirmation record. | The actual recipient or amount differs from what was authorized. |
| Provide stop and recovery paths. | Revocation test, incident record and manual fallback. | A revoked connector continues acting or retries a completed action. |
Three hypothetical product reviews
Internal drafting tool: assume access is genuinely limited to employees and it does not offer generative services to the domestic public. Record the Article 2 scope conclusion, but still investigate the data sent to vendors, permitted training use, retention and any overseas access. The non-public exclusion is not permission to upload client information without review.
Public image app with sharing: the generator, the content-sharing function and the app distributor need separate checks. A visible generator label does not prove that file metadata survives export, that the sharing service recognizes it or that app-review materials are complete. Use an end-to-end test file through generation, download, re-upload and public display.
Customer-service agent becoming a companion: adding persistent emotional memory, simulated intimacy or dependence-oriented engagement changes the facts behind the original scope analysis. Recheck the anthropomorphic rules and child safeguards as well as tool permissions. These are illustrative product reviews, not a statement that a real service is compliant.
Official sources & revisions
- 2023 Generative AI Interim Measures, Articles 2, 7–17, 19 and 22–24
- Algorithm Recommendation Provisions, Articles 2, 7–9, 16–17 and 24
- Deep Synthesis Provisions, Articles 2, 7–20 and 25
- 2025 AI Content Labeling Measures, Articles 2–12 and 14
- 2026 Anthropomorphic Interaction Measures, Articles 2 and 8–14
- Opinions on Regulated Application and Innovative Development of Intelligent Agents, items 4 and 6–11
: Expanded the 2023 service record into a bilingual rule-relationship study with role, date and evidence matrices. Distinguished labeling duties, filing triggers, emotional-interaction scope and non-binding agent policy; added original implementation scenarios.
Examples and working tables are original research aids, not official forms or legal advice.
Suggest a correction