Anthropomorphic AI Interaction: Draft and Final Rules Compared

Previous rule summary

The draft covered publicly offered products or services simulating human traits and engaging in emotional interaction, without an express task-service exclusion.

Compared rule summary

The final definition requires sustained emotional interaction. Listed task-oriented services without that feature are expressly outside these measures.

Practical impact: Assess actual functions, not the chatbot label. Exclusion from these measures is not exclusion from other AI or privacy rules.

Draft Article 2; final Article 2, including paragraph 3.

Previous rule summary

The draft already required lifecycle safety across design, operation, upgrades and termination, including monitoring, assessment, logs and dependence-risk safeguards.

Compared rule summary

The final retains lifecycle responsibility, lists deployment instead of design, and expressly adds privacy and personal-information protection to the required capabilities.

Practical impact: Retain lifecycle controls; do not label the entire duty newly introduced or discard design-stage work.

Draft Article 9; final Article 10. Corrects this site's earlier newness claim.

Previous rule summary

The emergency-response provision specifically required guardian or emergency-contact information at registration for minors and older users.

Compared rule summary

A separate final provision requires a service agreement, registration according to law and the agreement, and necessary information such as age, guardian or emergency contact.

Practical impact: Build a proportionate contact workflow. The examples do not mean every user must supply every field regardless of necessity.

Draft Article 11(3); final Article 12.

Previous rule summary

The draft expressly called for human takeover in specified self-harm or suicide situations, alongside guardian/emergency-contact notification.

Compared rule summary

The final requires timely supportive responses, necessary assistance/intervention in specified extreme situations and timely contact. It does not repeat the express human-takeover instruction.

Practical impact: Keep an operational escalation path. Omitting one prescribed technique does not remove intervention duties or prove unattended automation sufficient.

Draft Article 11; final Article 13.

Previous rule summary

The draft required guardian express consent for minors' emotional-companionship services, minors modes, controls and identity recognition.

Compared rule summary

The final prohibits virtual relatives, partners and other virtual intimate relationships for minors. Other covered services for children under 14 require guardian consent. Protective modes and controls remain.

Practical impact: Separate prohibited designs from services that may use a consent workflow. A parental checkbox cannot override the prohibition.

Draft Article 12; final Articles 14 and 17. Under-14 data processing has a separate consent requirement.

Previous rule summary

The draft prohibited simulating an older user's relatives or specified relationship persons and set out emergency-contact guidance.

Compared rule summary

The final older-user clause focuses on healthy-use guidance, prominent risk warnings, timely responses to help requests and protection of rights; the specific simulation ban is not repeated.

Practical impact: Do not present the omitted draft ban as current text. Its omission is not permission for deception, manipulation or violations of other rules.

Draft Article 13; final Articles 8, 13 and 15.

Previous rule summary

The draft required deletion options and security safeguards, including restrictions on third-party provision.

Compared rule summary

The final expressly adds copying as well as deletion options while retaining disclosure restrictions and security safeguards.

Practical impact: Provide both export/copy and deletion paths, identify affected stores and protect each request with identity and access checks.

Draft Article 14; final Article 16(1)-(3). Copying is the added element, not the whole protection regime.

Previous rule summary

The draft's special training restriction covered user interaction data and sensitive personal information, subject to a statutory exception or separate consent.

Compared rule summary

The final clause refers to interaction data that is the user's sensitive personal information, with an exception under laws/administrative regulations or separate consent.

Practical impact: Classify training inputs and record their legal basis. A narrower special clause does not allow unrestricted training on other identifiable chat records.

Draft Article 15(1); final Article 16(4). General privacy duties remain applicable.

Previous rule summary

The draft expressly specified an annual compliance audit when processing minors' personal information.

Compared rule summary

The final retains self-conducted or commissioned professional audits according to relevant national provisions, without repeating an annual frequency here.

Practical impact: Determine frequency under current applicable audit rules and facts, not the superseded draft. The audit duty itself was not removed.

Draft Article 15(2); final Article 17(2).

Previous rule summary

The draft required AI identity disclosure, first-use/re-login or dependence-triggered reminders, and a pop-up pause reminder after more than two continuous hours.

Compared rule summary

The final connects labeling and identity disclosure, retains dependence reminders, and requires dialogue or pop-up duration reminders each time continuous use exceeds two hours. First-use/re-login wording is not repeated.

Practical impact: Test recurring duration and risk-triggered prompts separately. This clause does not mandate a general automatic two-hour shutdown.

Draft Articles 16-17; final Article 18.

Previous rule summary

The draft already listed launch/additional functions, major technology changes, at least one million registered users or 100,000 monthly active users, and risk-related situations.

Compared rule summary

The final retains the main independent triggers, provides for report sharing through the provincial CAC, and expressly covers assessment required by a provincial-or-higher CAC notice.

Practical impact: Low user counts are not a launch exemption. The two user metrics are alternatives and do not replace the other triggers.

Draft Article 21; final Article 22. The numbers are retained, not new.

Previous rule summary

The draft required filing/change/cancellation, annual review of filing materials and provincial written review of reports and audits.

Compared rule summary

The final retains filing and annual verification. Provincial review refers to assessment reports and related circumstances; explicit audit references in that review and on-site audit wording are not repeated.

Practical impact: Keep filing, assessment and compliance-audit evidence distinct. Administrative verification is not the provider's own audit or a new blanket annual service licence.

Draft Articles 25-26; final Articles 26-27.

Previous rule summary

The draft required convenient exit and stopping on request, with notification/disclosure for withdrawal or unavailability.

Compared rule summary

The final retains exit rights, expressly includes voice controls and prohibits continued interaction obstructing exit. Cessation requires prior notice or, if impossible, a timely announcement.

Practical impact: Test exit commands and cessation messages; retention goals or emotionally persuasive replies cannot justify postponing exit.

Draft Articles 18-19; final Articles 19-20.

Previous rule summary

The draft referred first to existing laws/administrative regulations, then warnings, criticism, correction and suspension where those texts did not provide a rule.

Compared rule summary

The final adds fallback detail: registration/service restrictions; cessation for refusal to rectify or serious circumstances with a possible RMB 10,000-100,000 fine; an additional RMB 100,000-200,000 fine where life/health safety is endangered and harmful consequences occur.

Practical impact: Link consequences to their conditions and authority. These figures are not universal or maximum penalties for all AI/privacy violations.

Draft Article 29; final Article 30. Other applicable laws/administrative regulations are addressed first.