How Can I Localise My AI Solution for China?

Launching an AI product in China requires much more than translating its interface into Chinese.

An AI solution that performs well in other markets may become slow, inaccurate or unavailable when used from mainland China. Its model may not understand local terminology or user expectations. Its data flows may create compliance issues. Features that generate text, images, audio or video may also trigger requirements relating to algorithm registration, security assessment, content governance and AI-generated content labelling.

For foreign companies, the real question is therefore not simply, “How do we translate our AI product?” It is:

How do we adapt the product, model, infrastructure, data flows and operating processes so that the solution can work reliably and compliantly in China?

This article explains the main components of AI localisation for the Chinese market and outlines a practical route from initial assessment to launch.

What Does AI Localisation for China Actually Involve?

AI localisation is a combination of four workstreams:

1. Product localisation: adapting the interface, user journey, documentation and support for Chinese users.
2. Model localisation: improving the AI’s understanding of Chinese language, business context, terminology and acceptable outputs.
3. Technical localisation: deploying suitable infrastructure and integrations for reliable access within Mainland China.
4. Regulatory localisation: designing data handling, content controls, disclosures and filing procedures around Chinese requirements.

These workstreams are closely connected. For example, replacing an overseas model API with a China-hosted model may improve latency, but it can also change model performance, data processing responsibilities and the applicable filing route. The deployment model should therefore be decided before the product is fully rebuilt.

1. Choose the Right China Deployment Model

The first decision is where the AI service will run and which model it will use.

Option A: Use a China-based model and cloud environment

For many foreign companies, the most practical route is to deploy the China-facing product on mainland China cloud infrastructure and connect it to a suitable locally available model.

This can provide:

• Lower latency and more consistent availability for mainland China users
• Easier integration with local identity, payment, messaging and cloud services
• A clearer route for local data storage and operational control
• The possibility of using an already-filed generative AI model, subject to the application’s own registration and compliance requirements

This approach does not mean that the international product must be replaced. Companies can maintain a separate China architecture while sharing appropriate product logic, brand assets and non-sensitive content across markets.

AppInChina’s China LLM Index provides a regularly updated view of China’s leading LLM providers and AI assistants, ranked by monthly active users, with information on products including Doubao, Qianwen, DeepSeek, Yuanbao and Kimi. It can provide a useful starting point for understanding the market, although the final model should still be tested against the product’s own Chinese-language tasks and compliance requirements.

Option B: Continue calling an overseas model

Connecting Chinese users directly to an overseas AI service can appear simpler, but it may introduce several problems:

• Cross-border network latency and unstable service availability
• Restrictions affecting particular overseas services or dependencies
• Cross-border transfers of prompts, account information or usage data
• Difficulty providing local incident response, moderation and customer support
• Continued exposure to Chinese rules where the service is offered to the public in mainland China

China’s Interim Measures for Generative AI Services expressly address services provided from outside China to users inside the country. An offshore architecture should therefore not be treated as a way to avoid Chinese compliance obligations.

Option C: Use a private or hybrid deployment

For enterprise AI, regulated industries or products handling sensitive business information, a private or hybrid architecture may be more appropriate. A company might deploy a model in a dedicated China environment, use retrieval-augmented generation with an approved local knowledge base, and restrict selected workflows to authorised enterprise users.

The scope of China’s generative AI rules depends partly on how the service is offered. The Interim Measures apply to generative AI services offered to the public in mainland China, while research, internal enterprise applications and other uses not offered to the public fall outside those particular measures. Other requirements concerning personal information, cybersecurity, data security and sector regulation may still apply.

2. Localise the Model, Not Just the Interface

A fluent Chinese interface does not guarantee a useful Chinese AI experience. The model also needs to work with the language, knowledge and workflows of its target users.

Simplified Chinese and local communication patterns

The product should be tested for more than literal translation. Chinese users may use abbreviations, mixed Chinese and English terminology, industry-specific vocabulary, indirect requests and platform-specific expressions. Prompts, system instructions, error messages and safety notices should all be written and tested as native Chinese content.

Local knowledge and terminology

An international model may have limited or outdated knowledge of Chinese regulations, product categories, administrative regions, companies or technical standards. Depending on the use case, localisation may involve:

• A curated Chinese knowledge base
• Retrieval-augmented generation using approved local documents
• China-specific system prompts and response templates
• Fine-tuning or other model optimisation using lawfully sourced data
• Human review by Chinese-speaking subject-matter experts

Evaluation in realistic China scenarios

Benchmarking should reflect the tasks the product will actually perform. A customer-support assistant, for example, should be evaluated on Chinese product names, local refund policies, ambiguous questions, prohibited requests and escalation to a human agent, not only on generic language tests.

Testing should cover accuracy, hallucination, bias, refusal behaviour, privacy leakage, prompt injection, prohibited content and consistency across model updates. It should also include adversarial prompts written by native Chinese testers.

3. Design China-Specific Data Flows

AI products often process more data than a conventional application. Prompts may contain personal information, customer records, confidential documents or commercially sensitive material. Logs and model-improvement pipelines can create additional copies of the same information.

Before launch, the company should map:

• What information users submit to the AI
• Where prompts, outputs, embeddings and logs are stored
• Which entity decides how the information is processed
• Which model, cloud and analytics providers receive the information
• Whether any information is transferred outside mainland China
• Whether prompts or user interactions are used for training or optimisation
• How users can access, correct or delete their personal information

China’s Network Data Security Management Regulations require network data processors to implement security controls and set out personal-information transparency and user-rights requirements. China’s rules on cross-border data flows also establish different procedures depending on the type and volume of information exported. Some transfers may be exempt, while others can require a standard contract, certification or a government security assessment.

The answer is not always “store everything in China,” but local storage can reduce unnecessary cross-border transfers and make the architecture easier to operate. The correct solution depends on the data categories, processing purpose, user volume, sector and corporate structure.

4. Determine Which AI Rules Apply

China regulates AI through several overlapping rules. The obligations depend on the product’s functionality, users and distribution model.

Generative AI services

China’s Interim Measures for Generative AI Services apply when generative AI is used to provide text, images, audio, video or other generated content to the public in mainland China. Among other matters, providers are expected to:

• Use training data and foundation models from lawful sources
• Protect personal information and intellectual property
• Take measures to improve accuracy, reliability and non-discrimination
• Establish user agreements and complaint mechanisms
• Protect user prompts and usage records
• Address unlawful content and maintain appropriate records
• Complete applicable security assessment and algorithm filing procedures

The definition of a provider can include an organisation supplying generative AI through an API, not only the operator of a consumer chatbot.

Full AIGC Filing, AIGC Registration and Algorithm Filing

China distinguishes between the underlying generative AI service and an application that uses a model that has already completed filing.

• Full AIGC Filing is generally the relevant route when a company develops its own foundation model. The review can cover the model architecture, training-data provenance, safety alignment, generated-content testing, content controls and the live service. AppInChina’s guide, What Is China’s AIGC Filing?, explains the scope of the filing and the challenges foreign companies commonly face.

• AIGC Registration may be available where an app, website, SaaS platform, connected device or other product directly calls a model that has already completed China’s Generative AI Service Filing without independently modifying or deploying it. AppInChina’s Complete Guide to China’s AIGC Registration explains the eligibility criteria, required materials, product controls, local review process and the online launch number, or 上线编号, issued after registration.

According to the Cyberspace Administration of China, by 30 June 2026, 988 generative AI services had completed filing and 598 applications or functions using filed models had completed registration. Launched applications should display the relevant model name and filing or registration information in a prominent location or on the product information page.

Using a filed model can simplify part of the route to market, but it does not automatically make the whole product compliant. The application operator remains responsible for its own product design, data processing, disclosures, content controls and any applicable registration.

AI-generated content labelling

China’s Measures for Labelling AI-Generated and Synthetic Content took effect on 1 September 2025. Depending on the service and content type, providers may need to add:

• Explicit labels that users can clearly see or hear; and
• Implicit labels embedded in file metadata, including information identifying the generated nature and source of the content.

The requirements cover generated or synthetic text, images, audio, video and virtual scenes. Labelling should be built into the content generation, export and publishing workflow rather than added as a final design adjustment.

Deep synthesis and identity-related functions

Products that generate or modify faces, voices, video or highly realistic virtual scenes may also fall under China’s deep synthesis rules. These include obligations concerning real-name authentication for certain publishing functions, input and output review, security assessment, content labelling and user complaint channels.

If the product uses facial recognition to identify individuals, separate facial-recognition rules may also apply. The compliance analysis must distinguish between generating a face, editing a face and using facial data to verify or identify a person.

AI companion and emotional-interaction products

China’s Interim Measures for Anthropomorphic AI Interaction Services took effect on 15 July 2026. They apply to sustained emotional-interaction services that simulate human personality traits, thought patterns and communication styles. Ordinary customer service, knowledge Q&A, work assistants, education and research tools are excluded where they do not involve sustained emotional interaction.

Covered services face additional requirements relating to emotional dependency, crisis intervention, minors, interaction data, usage-time reminders, exit mechanisms, safety assessment and annual verification of algorithm filings. Companies developing AI companions, virtual partners or emotionally responsive characters should therefore treat this as a distinct product category.

AI products that also use blockchain

If the AI product also uses blockchain, the provider may need to complete a separate CAC blockchain information service filing and implement requirements relating to information-security responsibility, real-name authentication, service rules, content governance and user management. New blockchain products, applications or functions may also require a security assessment. AppInChina’s guide, What Is China’s Blockchain Information Service Filing?, explains when the filing applies and how it fits alongside the other approvals needed to operate a digital service in China.

5. Prepare the Product for Distribution in China

AI compliance is only one part of market access. A China-facing website or app may also require:

• A suitable Chinese entity or qualified local operating structure
• Mainland China cloud infrastructure and a Chinese cloud account
• ICP filing or, for certain commercial internet services, an applicable licence
• Mobile app filing
• China-specific privacy notices, user agreements and consent flows
• App-store submission materials and software copyright documentation where required
• Local SDK review, security testing and vulnerability remediation
• Local monitoring, moderation, complaint handling and incident response

The exact requirements depend on whether the product is a website, mobile app, mini program, API, embedded enterprise feature or another form of service. Sector-specific approvals may also apply in areas such as healthcare, finance, education, publishing and online audio-visual content.

A Practical Five-Stage Localisation Process

Stage 1: Product and regulatory assessment

Define the China use case, target users, AI functions, content types, distribution channels and commercial model. Identify which entity will operate the service and whether it will be public, enterprise-only or hybrid.

Stage 2: Architecture and model selection

Compare local model and cloud options against the product’s quality, latency, privacy and cost requirements. Map every data flow and decide which services must be deployed or replaced in China.

Stage 3: Product and model adaptation

Localise the interface, prompts, knowledge sources, safety rules and human-support processes. Integrate local services and test performance under real mainland China network conditions.

Stage 4: Compliance implementation

Prepare privacy and user documents, content governance procedures, AI labels, security materials and any necessary algorithm, generative AI, app or website filings. Compliance requirements should be reflected in the code and operating workflow, not only in legal documents.

Stage 5: Controlled launch and ongoing monitoring

Launch to a limited user group, measure output quality and safety, and refine the product before scaling. Continue monitoring changes to the model, training data, user volume and product functionality because a material change may require renewed testing, updated documentation or amended filings.

Common AI Localisation Mistakes

Foreign companies frequently encounter problems when they:

• Translate the interface but continue using an unreliable overseas technology stack
• Select a model before mapping the product’s data and compliance requirements
• Assume that using a filed Chinese model makes the entire application compliant
• Send all prompts and logs overseas without analysing cross-border data obligations
• Add content moderation only after model development is complete
• Treat AI-generated content labels as a manual publishing task
• Use international benchmarks instead of testing real Chinese users and edge cases
• Launch without a local process for complaints, incidents and regulator requests

The most effective localisation programmes address product, infrastructure and compliance together from the beginning.

How Can AppInChina Help?

AppInChina helps international companies assess, adapt and launch digital products for the Chinese market. For an AI solution, this can include:

• Evaluating the product’s China hosting, model and integration requirements
• Designing a reliable mainland China technical architecture
• Supporting Chinese cloud account setup and deployment
• Assessing ICP, app filing and app-store distribution requirements
• Coordinating localisation, SDK review and China-market testing
• Supporting the operational implementation of privacy, data and AI compliance requirements
• Managing launch and ongoing technical operations in China

There is no single architecture or filing checklist that works for every AI product. The correct route depends on what the AI does, who can use it, what data it processes and how it will be distributed.

If you are planning to launch an AI-enabled website, application or enterprise platform in China, contact us for a free initial assessment of your localisation, hosting and compliance requirements.