As AI customer service, digital-human hosts and enterprise agents become more human-like, regulation has followed.
On 10 April 2026 five central agencies — the CAC, the NDRC, the MIIT, the Ministry of Public Security and the State Administration for Market Regulation — jointly issued the Interim Measures for the Administration of Anthropomorphic Interactive AI Services (Order No. 21), effective 15 July 2026 (CAC official text; People's Daily Online, April 2026; Xinhuanet, 31 July 2026).
It is the world's first national-level regulation dedicated to anthropomorphic or "emotional AI" interaction (JT&N commentary; Just Security).
Almost simultaneously, QuestMobile's June ranking of domestic AI-native apps put Doubao at 382 million MAU, a clear first (Sina Finance, 5 August 2026). The larger the conversational entry, the more urgent compliance and visibility around "anthropomorphic" experiences become.
For enterprises, GEO is no longer only about being seen, but about being seen in a compliant way. If you operate any customer-facing AI touchpoint, a one-day compliance-and-visibility audit can map what needs labeling first — contact us to book it.
What the rules cover
The measures are China's first dedicated rules for anthropomorphic interactive AI. They target services that use AI to simulate a natural person's personality, thought patterns and communication style in sustained emotional interaction — typical scenes include AI companions, virtual partners and AI psychological-counseling products (CAC official text; EN Wikipedia).
Human-like image, voice and dialogue elements trigger identification, filing and security-assessment duties. Public reporting and industry readings point to five enterprise-relevant dimensions:
| Dimension | Core requirement | Direct impact |
|---|---|---|
| Identification | Mark AI interaction clearly so users are not misled | Human-like bots and digital-human content must disclose AI generation |
| Filing & security assessment | File anthropomorphic services and complete security assessment | In-house agents and AI CS must follow the compliance process |
| Content governance | Providers are responsible for interactive content and review | AI-generated assets need a compliance gate before entering the brand source library |
| Scope boundaries | General CS bots and knowledge Q&A are outside the strictest duties | Focus compliance energy on emotional / human-like touchpoints |
| Enforcement | Warnings, correction orders, suspension; fines of RMB 10,000–100,000 (up to 100,000–200,000 where life/health is harmed) | Violations are not just reputational — they carry regulatory cost |
The rules target anthropomorphic interaction, not AI search or GEO as such — but any outward human-like touchpoint (site chatbot, marketing digital human, social AI persona) needs matching compliance. Note the scope nuance: a general customer-service bot or knowledge Q&A that does not simulate sustained emotional interaction sits outside the strictest obligations (Vectrel, July 2026), while AI companions and human-like personas fall squarely inside.
We previously discussed trust in the AI era in EEAT for brands; the rules upgrade that from advice to requirement.
How identification affects AI search visibility
A less-discussed knock-on effect: engines will be more sensitive to "AI-generated" and "anthropomorphic" labels when retrieving and citing sources. Doubao, DeepSeek and Yuanbao already prefer verifiable sources (see AI engine source preferences). Mandatory labels add another authenticity signal — a page that honestly marks "AI-assisted" and backs claims with real, checkable facts scores better on the trust dimension than one that hides its AI identity.
Three GEO rules follow:
- Real first: bylined humans, real cases and verifiable data gain citation weight; anonymous marketing accounts and mass-generated fake personas lose it.
- Label as credit: disclosing "AI-assisted" is closer to engine preference than hiding AI identity — transparency is part of EEAT trust.
- Official-source dividend: filing and security assessment create a compliance backstop; structured content around filing facts can raise LLM inclusion (see Structured data and LLM inclusion).
Four steps for compliant GEO
- Inventory anthropomorphic touchpoints: site CS, digital humans, AI marketing assets — which need labels and filing, and which are ordinary functional tools.
- Register AI content: prominent labels, review and traceability, records you can show regulators.
- Strengthen a real source matrix: site, certificates, authoritative coverage and real cases, with entity consistency (name, address, contacts) so engines can find, recognize and cite you.
- Feed GEO with compliance content: turn filing, assessments and policies into FAQs and structured data — both an audit trail and Q&A material for Doubao and Yuanbao (see Doubao content optimization).
Compliance in a global context
China is not alone in moving. The EU AI Act's Article 50 chatbot-disclosure obligation applies from 2 August 2026; South Korea's AI Basic Law took effect on 22 January 2026; Japan passed an AI promotion law in May 2025 using a voluntary "soft law" approach; and US states such as California and New York have companion-AI disclosure statutes (JT&N; Vectrel; Just Security).
For multinational brands this means one principle — disclose human-like AI, keep real sources behind it — is becoming a global GEO baseline, not a China-only constraint.
FAQ
Must all AI content be labeled? No. The measures target anthropomorphic interactive services — human-like image, voice or dialogue with sustained emotional interaction. Ordinary AI search Q&A, knowledge tools and general customer-service bots are out of scope, but companies offering AI companions, digital humans or emotional personas should self-audit.
Does GEO itself cross a compliance line? No. GEO raises citation probability through real, authoritative, structured sources — aligned with "real and transparent." Risk sits with fake personas and impersonating humans.
Do enterprise agents need filing? Anthropomorphic interactive services should follow filing and security assessment; exact boundaries follow implementing rules. General work-assistant agents that do not simulate sustained emotional interaction are lighter in scope. Seek compliance advice before deploying agents, and put filing into the GEO timeline.
Will the rules change how Doubao includes sources? Engines will keep raising weight on authenticity and verifiability. Doubao already has 382 million MAU (QuestMobile June ranking). Compliance labels become a new filter — making official and third-party sources solid is the safest hedge.
Is this the first law of its kind globally? Yes, China's interim measures are the first national-level regulation dedicated to anthropomorphic or emotional AI interaction; other regions rely on general AI laws, EU chatbot-disclosure rules or US state statutes (JT&N; Just Security).
What penalties can providers face? Warnings, criticism notices, correction orders or suspension; fines of RMB 10,000–100,000, rising to 100,000–200,000 where harm to life or health results (Article 30, CAC official text).
Where should SMEs start with limited budget? Label site CS and AI marketing first (lowest cost), then entity consistency and site FAQs, then filing and authoritative coverage as needed. Compliance and visibility rise together.
Related reading
This article was written by Zheming Digital Communication Research Institute. Data updated to 2026; based on public reporting: CAC official text (Order No. 21, effective 15 July 2026), People's Daily Online (13 April 2026), Xinhuanet (31 July 2026), JT&N law-firm commentary, Just Security and Vectrel analyses (2026), and QuestMobile June AI-native MAU ranking via Sina Finance (5 August 2026). Readings are industry observation; operations must follow official texts.
Shanghai Zheming provides GEO, Doubao content optimization, AI visibility diagnosis and website development so brands can be seen by AI under the new rules. +86 18917757529 · jaysun@widesight.cn · contact us.