If you are an AI company planning to enter the Chinese market, algorithm filing (算法备案, suànfǎ bèi’àn) is almost certainly on your to-do list. It is one of the most important — and most frequently misunderstood — regulatory requirements for AI products in China.

This guide explains what algorithm filing actually is, who needs it, how the process works, how long it takes, and what it costs. Written specifically for foreign companies entering China.


What Is Algorithm Filing in China?

China Algorithm Filing: A Complete Guide for Foreign Companies (2026) - AI Law & Regulation infographic

Algorithm filing is a regulatory requirement under China’s Regulations on the Administration of Internet Information Service Algorithmic Recommendations (互联网信息服务算法推荐管理规定), which took effect on March 1, 2022.

In plain English: If your service uses algorithms to recommend content or information to Chinese users, you must register the algorithm with the Cyberspace Administration of China (CAC).

There are actually two separate but related filing regimes:

RequirementChinese NameApplies To
Algorithm filing算法备案Services using recommendation algorithms (news feeds, product recommendations, etc.)
Generative AI service registration生成式人工智能服务备案Services generating text, images, audio, video for the public (LLMs, AI chatbots, image generators, etc.)

Many AI products need both. A chatbot powered by a large language model, for example, requires generative AI service registration and may also require algorithm filing if it has recommendation features.


Who Needs to File?

If you answer “yes” to any of these, you likely have filing obligations:

What about foreign companies?

This is the question we get most often. The short answer: if your service reaches users in mainland China, the rules apply to you — whether you have a local entity or not.

In practice, however, the filing system is designed around domestic entities. There are three common scenarios for foreign companies:

Scenario 1: You have a Chinese entity (WFOE or JV)
You file through your domestic entity. This is the standard and most straightforward path.

Scenario 2: You do not have a Chinese entity but serve Chinese users
You still have compliance obligations, but the filing mechanism is less clear. In practice, most regulators expect you to work through a local partner or set up a domestic entity.

Scenario 3: You only sell into China through an API or B2B model
If your customers are businesses rather than end users, and you do not directly provide services to the Chinese public, the filing obligations may fall on your customer rather than you. But this depends on the specific arrangement and should be assessed carefully.

Key takeaway: For most foreign AI companies planning a serious China entry, setting up a domestic entity is the practical starting point for algorithm compliance.


The Filing Process Step by Step

Step 1: Assess which filings you need

The first step is to map your product to the specific requirements. We typically look at:

Step 2: Prepare filing materials

The documents required depend on the filing type, but typically include:

Step 3: Submit through the filing system

Filing is done through the CAC’s online Internet Algorithm Filing System (互联网信息服务算法备案系统). The process is in Chinese and requires a Chinese entity account.

Step 4: Review and revisions

Regulators review the submission and may ask for revisions or additional information. This is normal — first-pass approval rates are low, especially for complex AI products.

Step 5: Approval and display

Once approved, you receive a filing number which must be displayed on your service (usually in the settings or about section). You then have ongoing obligations to update filings when the algorithm changes significantly.


Timelines

StageAlgorithm FilingGenerative AI Registration
Material preparation2–4 weeks4–8 weeks
Initial review10–20 working days20–30 working days
Revisions (if needed)1–2 rounds2–3 rounds
Safety assessmentNot required30–60 working days
Total (typical)4–8 weeks3–5 months

These are rough estimates. The actual timeline depends heavily on:


Costs

Filing itself has no government fee. The cost comes from legal and technical work:

For foreign companies, there is also the cost of setting up a domestic entity if you do not already have one.

We provide fixed-scope packages for algorithm filing so you know the total cost upfront. Contact us for a quote based on your specific product.


Common Mistakes Foreign Companies Make

1. Assuming “we are overseas, so the rules don’t apply”

This is the most common — and most risky — misconception. China’s algorithm regulations apply to services provided to Chinese users, regardless of where the company is based. Enforcement against overseas services is increasing.

2. Confusing algorithm filing with generative AI registration

They are two different requirements with different processes and timelines. Many products need both. Starting with only one and discovering the other later can delay your launch by months.

3. Trying to file before the product is ready

You do need a working product to file — regulators expect to review an actual service. But you should plan the compliance architecture before building, because product design decisions directly affect what you need to file.

4. Underestimating the documentation work

The filing system is thorough. You need detailed algorithm descriptions, data source documentation, content governance procedures, user rights mechanisms, and more. It is not a simple form — it is a substantial compliance document package.

5. Filing and then forgetting about it

Algorithm filing is not one-and-done. You have ongoing obligations:


How We Help Foreign Companies

We work with AI companies from the US, EU, Singapore and beyond to handle their China algorithm compliance. Our process:

  1. Assessment — We review your product and tell you exactly what filings you need (and what you don’t)
  2. Roadmap — We give you a clear timeline, cost estimate, and action plan
  3. Preparation — We prepare all filing materials in both English and Chinese
  4. Submission — We handle the filing process and all regulator communications
  5. Ongoing support — We keep you updated on regulatory changes and handle ongoing compliance

You work with the same team from first assessment to long-term compliance — no handoffs, no information lost in translation.


Next Steps

If you are planning to enter the Chinese market with an AI product, the best time to start thinking about algorithm compliance is now — ideally before you have built your China product architecture.

Start an Enquiry — tell us about your product and China plans, and we will give you a clear assessment of what you need, how long it will take, and what it will cost.

Multi-Jurisdiction AI Compliance: China, EU and US Side by Side

Foreign AI companies often operate across multiple jurisdictions simultaneously, and the compliance burden multiplies quickly. China’s algorithm filing regime, the EU AI Act, and US state-level AI laws each have different philosophies, requirements and enforcement styles.

The EU AI Act follows a risk-based classification model — prohibited, high-risk, medium-risk, and limited-risk — with corresponding obligations. China’s regime is more filing-centric, with specific registration requirements for different algorithm categories. The US approach remains patchwork, with state-level laws (like California’s AI transparency rules) and sector-specific federal guidance rather than a single national framework.

For companies subject to multiple regimes, the practical strategy is to identify the areas of overlap and build a single compliance infrastructure that satisfies all of them. Data mapping, for example, is a requirement under China’s PIPL, the EU GDPR, and several US state laws. Building one robust data map that is structured to answer each jurisdiction’s questions is far more efficient than running three separate compliance projects.

Looking at CAC enforcement actions from 2025 and the first half of 2026 reveals clear patterns about where regulators are focusing their attention. The most common violations are not grand failures — they are procedural gaps: incomplete filing descriptions, unreported algorithm modifications, and failure to display filing numbers to users.

Penalty amounts vary widely, from RMB 100,000 to several million yuan, depending on whether the violation is considered a first-time oversight or a deliberate circumvention. What is consistent across cases is that regulators expect companies to maintain their filings actively, not just submit them once and forget.

For foreign companies, the enforcement pattern has a specific implication: your compliance team needs a calendar. Algorithm change reports, annual filing reviews, and deep-synthesis label audits all have deadlines that, if missed, compound quickly. A structured compliance calendar with advance reminders is the single most cost-effective way to stay on the right side of enforcement.

How AI Regulation Maps to Your China Entry Timeline

Foreign companies often ask when in their China entry process they should start planning for AI regulation. The short answer: before you begin product localisation. Algorithm filing is not a last-minute compliance checkbox — it requires scope analysis, document preparation, and iterative feedback with the Cyberspace Administration of China (CAC) that typically takes four to eight weeks. If your launch timeline is tight, you should start the filing assessment during the product design phase, not after you have a working demo.

The regulatory landscape also interacts with your entity structure. A wholly foreign-owned enterprise (WFOE) that provides AI services directly to Chinese users has different filing obligations than a representative office or a cross-border service model. Getting the entity structure right first can simplify — or complicate — every subsequent compliance step.

For generative AI specifically, the registration process has three gates: pre-launch security assessment, formal filing, and ongoing supervision. Companies that plan for all three from day one tend to launch faster than those that treat filing as a single, isolated task.

The Difference Between Filing, Registration and Security Assessment

One of the most common points of confusion for foreign companies is the relationship between algorithm filing (算法备案), generative AI service registration (生成式人工智能服务备案), and security assessment (安全评估). These are three distinct processes with different triggers, timelines and regulators.

Algorithm filing is the broadest requirement: it applies to any algorithmic recommendation system, deep synthesis service, or personalised push system with public-facing users. Generative AI service registration is a subset — it applies specifically to large language model services and requires additional steps, including a pre-launch security assessment. The security assessment itself is the deepest review, examining training data sources, content generation mechanisms, and safety guardrails.

Practically, this means a foreign company launching a generative AI product in China will likely go through all three: first the security assessment (the gate), then the generative AI service registration (the licence equivalent), and the algorithm filing (the ongoing registry). Understanding this stack before you start avoids the common mistake of preparing for one process only to discover another is also required.

Regulatory Arbitrage and Why It Usually Fails

Some foreign companies consider structuring their China operations in ways that try to minimise or avoid regulatory obligations — for example, by keeping all servers overseas, using a Hong Kong entity instead of a mainland China entity, or working through a local partner to avoid direct filing requirements. These approaches rarely work as well as hoped.

Chinese regulators have been increasingly willing to look through corporate structures to the substance of what is actually happening. If your service is targeting Chinese users, if it processes their data, or if it has a real connection to the China market, regulators will generally find that Chinese rules apply — regardless of where your servers are located or what entity technically operates the service.

This is not to say that structure does not matter — entity structure, data architecture, and operating model all have significant implications for exactly which rules apply and how you comply. But trying to avoid regulation entirely through structuring is usually a losing game. The better approach is to understand the rules, design your structure to comply efficiently, and build compliance into your operations from the start.

Preparing for Regulator Communication: What to Expect

Foreign companies are sometimes surprised by how interactive China’s AI regulatory process is. Unlike jurisdictions where you submit a form and wait for a yes/no answer, algorithm filing and generative AI registration often involve multiple rounds of questions, clarifications, and adjustments.

The regulator’s questions typically fall into a few categories: scope clarification (is your product correctly categorised?), data sourcing (where does your training data come from?), safety mechanisms (how do you prevent harmful outputs?), and user protection (what disclosures and controls do users have?). Each round generally takes one to two weeks, and there are typically two to four rounds for a straightforward case.

The key to smooth regulator communication is preparation. Before you file, you should have clear, documented answers to the questions you are most likely to receive. This not only speeds up the process but also builds confidence with the regulator — a factor that matters more than many foreign companies realise.

The Role of Industry Associations and Standards Bodies

Navigating AI regulation in China is not just about interacting with government agencies. Industry associations and standards bodies play a significant role in shaping regulations, interpreting requirements, and providing guidance to companies. Understanding this ecosystem can help foreign companies stay ahead of the curve.

Key bodies include the China Artificial Intelligence Industry Development Alliance (AIIA), the China Electronics Standardization Institute (CESI), and the China Internet Network Information Center (CNNIC). These organisations develop technical standards, issue implementation guidelines, and sometimes run pilot programmes that give participating companies early insight into regulatory direction.

For foreign companies, engaging with these bodies can provide several benefits: early access to draft standards, opportunities to provide feedback, networking with industry peers and regulators, and credibility with both regulators and customers. Participation does not require a China entity in all cases — international companies can often join as observers or associate members. The level of investment you should make depends on how central China is to your business and how much regulatory risk you face.

How to Build an In-House Compliance Team in China

As your China operations grow, you will eventually need to bring some compliance functions in-house rather than relying entirely on external advisors. The question is: when, and what roles do you need?

For early-stage companies with a small China team, it usually makes sense to outsource most compliance work to external counsel and consultants. As you grow to 20-30 China employees and have a launched product, you should consider hiring your first compliance person — typically a senior compliance manager or data protection officer who can handle day-to-day compliance coordination and work with external advisors on more complex issues.

As you scale further — say 50+ employees and multiple products — you may need a small team: a compliance lead, someone focused on data protection, and someone focused on regulatory filing and government relations. The exact structure depends on your industry, your regulatory risk profile, and your growth trajectory. The key is to grow the team proactively — before a crisis or a regulatory gap forces you to — but not so fast that you are paying for capacity you do not need.

Annual Compliance Planning for AI Companies in China

AI regulation in China is not a one-time project — it is an ongoing obligation. Companies that treat it as a checkbox exercise consistently run into problems. Those that build structured annual compliance programmes, on the other hand, find that the burden decreases over time as systems mature.

A well-structured annual compliance calendar should include: quarterly reviews of algorithm changes to identify filing update triggers, an annual filing review and renewal process, periodic deep synthesis label audits, regular training for product and engineering teams, an annual data protection impact assessment for high-risk processing, and monitoring of regulatory developments with impact assessments.

The rhythm matters. Quarterly check-ins keep compliance on the radar without overwhelming the team. Annual reviews provide an opportunity to step back and assess whether the overall programme is still fit for purpose. And regular training ensures that new team members understand their compliance responsibilities and that existing knowledge stays current as regulations evolve.

AI Agent Regulation: The Next Frontier

As AI agent technology advances from simple chatbots to autonomous systems that take actions on behalf of users, regulators are grappling with how to apply existing frameworks to these new capabilities. China has been at the forefront of this conversation, with early guidance and regulatory exploration already underway.

The key regulatory questions for AI agents include: Who is responsible when an agent makes a mistake or causes harm? How do transparency and disclosure requirements apply to systems that act semi-autonomously? What level of human oversight is required? And how do existing algorithm filing requirements map to agent architectures?

For companies building AI agent products, the practical advice is to design with regulation in mind, even where specific rules do not yet exist. Build in human oversight mechanisms, maintain clear audit trails of agent actions, provide transparent disclosures to users about what the agent can and cannot do, and design accountability structures so you know who is responsible when something goes wrong. The companies that get this right now will be well positioned when more specific agent regulations emerge.

FAQ

Q: How long does algorithm filing take in China?

A: For a straightforward case, algorithm filing typically takes four to eight weeks from submission to approval. Generative AI service registration takes longer — usually two to four months — because it includes a security assessment step. The timeline can vary depending on the complexity of your product and how many rounds of questions you receive from the regulator.

Q: Do foreign companies need a China entity to file algorithms?

A: Yes, generally you need a China-registered entity (typically a WFOE) to be the filing entity. The algorithm filing system is designed for entities operating within China. If you do not yet have a China entity but are planning to offer AI services to Chinese users, you should plan the entity setup and filing process together.

Q: What happens if we don’t file and just launch anyway?

A: Operating an unfiled algorithmic service can result in fines, orders to rectify, and in serious cases, suspension of service. The CAC has been actively enforcing filing requirements since 2023, and enforcement has intensified in 2025-2026. For foreign companies specifically, an enforcement action can also damage your reputation and complicate future regulatory interactions.

Q: Can we use our overseas LLM in China without local deployment?

A: It depends on how the service is structured. If Chinese users access your overseas-hosted model directly from outside China, the analysis is different from a locally deployed service. However, the regulatory trend is toward extending oversight to services that target Chinese users regardless of where the servers are located. The specific facts matter a lot — you should get legal advice based on your exact architecture and user base.

Q: How often do we need to update our filing?

A: You need to update your filing within ten working days of any material change to your algorithm — including changes to functionality, data sources, or recommendation logic. You also need to review your filing annually. Many companies set up a quarterly compliance review to catch changes that might trigger a filing update before the deadline becomes urgent.

Frequently Asked Questions↑ Back to article

Do foreign companies need to file algorithms in China?

If your company provides algorithm-recommendation services or generative AI services to users in mainland China — even through a cross-border service model — you likely have filing obligations. The specific path depends on whether you have a domestic entity, what type of service you provide, and how your algorithm is deployed.

What is the difference between algorithm filing and LLM registration?

Algorithm filing (算法备案) applies to recommendation algorithms and is the baseline requirement. Generative AI service registration (生成式人工智能服务备案) is an additional requirement for services that generate text, images, audio or video for the public. Many AI products require both. Registration is more rigorous and includes a safety assessment.

How long does algorithm filing take in China?

Algorithm filing typically takes 20–40 working days from submission to approval. Generative AI service registration takes longer — typically 60–90 working days including the safety assessment period. The actual timeline depends on whether your materials are complete on first submission and whether regulators ask for revisions.

Can a foreign company file without a Chinese entity?

In most cases, a domestic entity is required to serve as the filing entity. There are limited exceptions for certain cross-border scenarios, but these are narrow and require specific approvals. For most foreign AI companies entering China, setting up a WFOE or partnering with a local entity is the practical first step.

What documents are required for algorithm filing in China?

Required documents include entity qualification materials such as a business licence and ICP filing proof, an algorithm description explaining what the algorithm does and how it works, an algorithm mechanism explanation detailing how recommendations or generation function, user rights protection measures showing how users can opt out or adjust settings, content governance rules describing how harmful or illegal content is prevented, and data security measures. For generative AI filings, additional materials include training data sources, a safety assessment report, and a content labelling scheme.

What is the step-by-step process for algorithm filing in China?

The filing process has five main steps: first, assess which filings you need by mapping your product against the regulatory requirements; second, prepare all filing materials including algorithm descriptions, governance rules, and security measures; third, submit through the CAC's online Internet Algorithm Filing System using a Chinese entity account; fourth, respond to regulator review and any requested revisions, which is a normal part of the process since first-pass approval rates are low; and fifth, once approved, display the filing number on your service and maintain ongoing obligations to update filings when material changes occur.

What are the ongoing obligations after algorithm filing approval?

After approval, you must display your filing number on the service, typically in the settings or about section. You must update your filing within ten working days of material changes to the algorithm, and conduct annual reviews of your compliance. As regulations evolve, you may need to adjust your filing to meet new requirements. Building compliance into your product development process is more efficient than treating filing as a one-time project, since algorithm changes are frequent in AI products.

Geason Law Offices AI Industry Research Group | Reviewed by He Qianggao (Lawyer License No. 14403201310092337, 10+ years of practice)