By Yoel Molina, Esq., Owner and Operator of the Law Office of Yoel Molina, P.A.
China has built one of the world’s most comprehensive—and prescriptive—AI governance frameworks. Instead of a single “AI Act,” Beijing layers rules across content governance, algorithms, data, and cybersecurity. In 2025, lawmakers added fresh obligations for AI content labeling and moved to update the landmark Cybersecurity Law to explicitly cover AI safety and oversight. Below is a practical, source-backed explainer of China’s model and how it’s evolving.
The four legal pillars behind China’s AI rules
1) Cybersecurity Law (CSL). The CSL is China’s foundational digital statute (effective 2017). In late October 2025, the National People’s Congress adopted amendments that add explicit articles on the “safe and sound development of AI,” with the revision to take effect on January 1, 2026. Expect strengthened oversight, clearer incident-reporting, and alignment with AI-specific measures. (
DigiChina)
2) Data Security Law (DSL) and Personal Information Protection Law (PIPL). Together these laws govern data classification, “important data,” security obligations, and individual privacy rights. They apply broadly to data handling inside China and sometimes extraterritorially when Chinese residents’ data is processed. In March 2024, the Cyberspace Administration of China (CAC) issued
Provisions on Promoting and Regulating Cross-Border Data Flows, creating carve-outs and easing some transfer burdens (e.g., when data isn’t designated “important”). (
DigiChina)
3) Algorithmic & deep synthesis rules. China regulates the
use of algorithmic recommender systems (think feeds, rankings, personalization) and the
creation/distribution of synthetic media (voice, image, video).
-
Algorithmic Recommendation Provisions (effective March 2022) require filing, transparency, a “turn off recommendations” option, and content controls. (
DigiChina)
-
Deep Synthesis Provisions (effective January 2023) mandate watermarking/labeling, consent for using a person’s likeness/voice, and security assessments for certain services. (
chinalawtranslate.com)
4) Generative AI service rules. The
Interim Measures for the Management of Generative AI Services (effective August 2023) apply to public-facing GenAI services in mainland China. They require safety assessments, dataset governance (e.g., lawful sourcing), content moderation aligned with “core socialist values,” and mechanisms to handle complaints and rectify illegal content. (
chinalawtranslate.com)
2025: Two important additions—content labeling and a Cybersecurity Law refresh
Nationwide labeling of AI-generated content. In March 2025, Chinese regulators issued
Measures for Labeling AI-Generated Content (in force September 1, 2025). Labeling must be explicit (e.g., visible watermarks/captions) and
implicit (machine-readable metadata), and platforms must enforce compliance. National standards bodies also published a companion technical standard on labeling methods. (
Reuters)
CSL amendment adds AI oversight. At the end of October 2025, the NPC Standing Committee approved amendments to the Cybersecurity Law adding provisions on AI development and safety; official outlets describe enhanced monitoring and regulation of AI risks, with the update effective January 1, 2026. This will likely tighten incident response and clarify regulators’ authority over AI infrastructure and services. (
en.npc.gov.cn.cdurl.cn)
How China’s AI rules actually work in practice
A “content-first” compliance lens. Most AI obligations sit inside China’s information services system. Public-facing models must align outputs with approved content rules, implement watermarking/labeling, and build complaint and takedown channels. If a system can shape public opinion (recommendation feeds, chatbots), it triggers filing and additional transparency duties. (
DigiChina)
Security assessments and filings. Depending on scope and sensitivity, AI providers may need to pass CAC-led security assessments and register algorithms. Filing typically covers the model’s purpose, training data sources, and safety controls; regulators can order rectifications. (
DigiChina)
Data controls and cross-border transfers. PIPL/DSL/CSL form the “data triangle.” For firms exporting data (e.g., cloud analytics, customer support), the March 2024 cross-border rules carved out scenarios that avoid a full CAC security review, but “important data” remains tightly controlled and can be defined by sectoral catalogs or regulator notice. (
chinalawtranslate.com)
Labeling and provenance. Under the 2025 measures and earlier deep synthesis rules, AI-generated or manipulated content must be clearly marked for users and identifiable to platforms and crawlers—an attempt to deter misinformation, fraud, and impersonation. (
Reuters)
Standards + enforcement. China often pairs rules with national standards (GB/GB-T). The new labeling regime sits alongside a national standard defining labeling methods. Expect standards to inform product testing, platform audits, and enforcement campaigns. (
Global Practice Guides)
What this means if you operate in or with China
1) If you offer public GenAI or creative tools in China:
-
Build
content filters aligned with Chinese content categories and “core socialist values,” plus user reporting and appeal channels.
-
Implement
dual labeling (visible + metadata) and keep logs proving labels persist across edits and reposts.
-
Prepare for
algorithm filing and
security assessment where services have scale or social impact. (
chinalawtranslate.com)
2) If you run private/enterprise AI on China data:
3) If you’re a global brand repurposing China-origin content or models abroad:
4) If you’re a platform or marketplace:
China vs. EU/US: Key differences in approach
-
Regulatory home. China roots AI governance in information control (CAC-led), cybersecurity, and data laws; the EU centralizes obligations in a horizontal
AI Act; the U.S. is a patchwork (sectoral + state AGs + FTC). (
DigiChina)
-
Outputs over inputs. China’s rules emphasize
what AI says or shows (label, watermark, align with content rules) even more than model-risk tiers. The EU prioritizes
use-case risk; the U.S. emphasizes unfair/deceptive practices and civil rights in context. (
chinalawtranslate.com)
-
Data transfers. China’s 2024 provisions softened—but did not erase—cross-border frictions. Many firms still perform transfer assessments and segment systems to avoid handling “important data.” (
chinalawtranslate.com)
-
Rapid standardization. China publishes national standards that can become de facto audit checklists (e.g., labeling methods), speeding operational enforcement. (
Global Practice Guides)
A practical compliance checklist for teams interacting with China
Frequently asked questions
Do the 2025 labeling rules apply to B2B tools? If content can be shared to the public internet (directly or via your platform), labeling obligations likely attach. Internal-only generation in a closed environment is lower risk, but platforms still need provenance metadata and controls for any external export. (
chinalawtranslate.com)
What if our GenAI product is only available outside mainland China but used by Chinese tourists or travelers? Territoriality and targeting matter. If you
provide services to the public in mainland China, the GenAI measures can apply. When in doubt, geo-fence access and avoid local marketing until your compliance posture is ready. (
chinalawtranslate.com)
Has China relaxed data exports? Somewhat. The March 2024 provisions create exemptions and assumptions (e.g., you’re not handling “important data” unless a catalog or notice says so). But sensitive sectors and large volumes can still trigger assessment or contract routes. (
chinalawtranslate.com)
Contact Us
If you’re evaluating AI offerings that touch China—product launches, content platforms, data transfers, or vendor diligence—contact Attorney Yoel Molina at
admin@molawoffice.com, call
(305) 548-5020 (Option 1), or message via
WhatsApp at (305) 349-3637.