As large model capabilities enter enterprise business through APIs, SaaS platforms and agent tools, more and more enterprises are adding generative artificial intelligence features to scenarios such as customer service, marketing, knowledge management, education and training, and office collaboration. One question that is often underestimated is: after an enterprise connects to a large model, can it simply go live?

Many enterprises believe that because they have not trained a foundation model themselves and merely call the API of a model vendor, compliance responsibility should rest with the model vendor; other enterprises, seeing that the underlying model they use has already completed filing, believe that their own products need not go through any further procedures. Current regulatory rules and regulatory practice show that these understandings are not accurate.

Generative artificial intelligence regulation does not only cover enterprises that "build models". For enterprises that embed generative AI capabilities into applications, mini-programs, websites, agents or business systems and provide services to the domestic public, obligations may also be triggered separately, including generative artificial intelligence service filing, generative artificial intelligence application or function registration, algorithm filing and security assessment.

What enterprises really need to answer is not only "which model is used", but also: to whom the service is provided, in what manner the service is provided, whether an already-filed model is called directly, and whether substantive training or modification has been carried out on the model.

01. AI filing is no longer a "paper system" but has truly entered the enforcement stage

An enterprise launched an AI application.

The enterprise did not train its own large model; instead, it directly called the API of a large model that had already completed filing and integrated AI Q&A functionality into its own product. The enterprise believed that since the underlying model had already been filed and its own product merely called the interface, it naturally did not need to go through filing or registration procedures.

However, in a special inspection of generative artificial intelligence conducted by the local cyberspace administration, this understanding was not accepted. The regulator held that the product already constituted an application providing generative artificial intelligence services to the public, and that it should perform the corresponding filing or registration procedures in accordance with current regulatory requirements and complete rectification in accordance with the law.

Such regulatory practice is not an isolated case. In April 2026, the Cyberspace Administration of China deployed the "Qinglang · Rectifying Chaos in AI Applications" special action, listing "failure to perform large model filing and registration obligations as required" as key rectification content in the first phase, and treating issues such as "should be filed but not filed" and "should be registered but not registered" as targets of the special governance. Nationwide regulation began to form a new pattern combining "local enforcement + special rectification". Since the "Qinglang · Rectifying Chaos in AI Applications" special action was launched in April 2026, the first phase has cumulatively handled more than 14,000 AI products such as non-compliant websites, applications and agents, cleaned up more than 6 million pieces of illegal and non-compliant information, handled more than 26,000 accounts, and removed more than 1,300 non-compliant AI products from shelves. ①

In its successive filing announcements, the Cyberspace Administration of China has repeatedly emphasized:

"For generative artificial intelligence applications or functions that directly call the capabilities of already-filed models through API interfaces or other means, registration shall be carried out by local cyberspace administration departments."

As of April 30, 2026, 868 generative artificial intelligence services nationwide had completed filing, and 530 generative artificial intelligence applications or functions had completed registration; filing and registration have formed two regulatory paths running in parallel. ②

It is precisely against this regulatory background that a question almost every enterprise encounters has begun to arise:

If an enterprise merely calls an already-filed large model API, does it still need to go through filing? When is registration required? When is neither filing nor registration required?

This article hopes to systematically sort out the above questions in light of current regulatory rules, the latest enforcement practice and common enterprise application scenarios.

02. China's generative artificial intelligence regulation is not a single filing but a multi-layered structure

Generative AI regulation is not constituted by a single regulation or a single procedure; rather, it is built on the basic legal framework of cybersecurity, data security and personal information protection, and is superimposed with AI-specific rules, pre-launch procedural obligations and post-launch continuous compliance obligations.

At the bottom layer are basic laws such as the Cybersecurity Law, the Data Security Law and the Personal Information Protection Law. As long as an AI application processes personal information, calls enterprise data, uploads business materials or provides internet information services to users, the enterprise still needs to determine the lawful basis for personal information processing, the rules on sensitive personal information, data security measures and network operation security responsibilities.

Above the basic laws are special regulatory rules for algorithms and generative AI, including the Provisions on the Administration of Algorithm Recommendations for Internet Information Services, the Provisions on the Administration of Deep Synthesis of Internet Information Services, the Interim Measures for the Administration of Generative Artificial Intelligence Services and the rules on labeling AI-generated synthetic content, etc. ③④

Further up are the procedural obligations most easily confused by enterprises, including generative artificial intelligence service filing, generative artificial intelligence application or function registration, algorithm filing and security assessment under specific circumstances. At the top layer are continuous obligations after the product goes live, such as prevention of illegal and harmful information, labeling of generated content, personal information protection, complaint and report mechanisms, security incident handling and public disclosure of filing and registration information. Filing or registration does not mean that compliance work is over; it is merely a pre-procedure for the product to enter standardized operation.

The Interim Measures for the Administration of Generative Artificial Intelligence Services apply to services that use generative artificial intelligence technology to provide the domestic public of the People's Republic of China with content such as text, images, audio and video; research and development and application activities that do not provide generative artificial intelligence services to the domestic public do not apply to these Measures. ⑤

This provision constitutes the first entry point for enterprises to determine filing and registration issues: whether generative AI services are provided to the domestic public of China.

03. First distinguish whether the enterprise is an AI "user" or an AI service "provider"

Enterprises use AI in many ways, but from the perspective of filing and registration, scenarios can first be divided into two major categories according to whether services are provided externally.

Enterprises that are merely internal users are usually not subject to filing or registration

Where enterprise employees use already-launched general large models, domestic model platforms or AI features in office software to complete copywriting, translation and information organization, this usually constitutes use of AI services already provided by external service providers. In this relationship, the enterprise is mainly a user, not a provider of generative AI services to the public.

Similarly, where an enterprise deploys a knowledge Q&A tool in a closed internal environment and opens it only to internal employees, in principle this also does not constitute providing services to the domestic public as referred to in the Interim Measures for the Administration of Generative Artificial Intelligence Services.

However, "no need for generative AI filing or registration" does not mean there are no compliance requirements. When employees upload customer materials, contracts, resumes, trade secrets or sensitive personal information, the enterprise still needs to address issues such as data use permissions, third-party processing, personal information notice and consent, trade secret protection and cross-border transmission.

When an enterprise embeds AI into its own products, its identity may change from user to provider

When an enterprise adds AI capabilities to its own App, website, mini-program, agent or SaaS product and directly provides users with functions that generate text, images, audio or video, the enterprise is no longer merely an ordinary user purchasing AI tools.

At this point, even if the underlying inference capability comes from a third-party model, the object that users contact, use and establish a service relationship with may still be the enterprise's product. The enterprise needs to determine whether its role has changed into that of a provider of generative AI applications or services.

Regulation needs to know not only who provides the underlying model, but also which applications are providing services to the public through that model.

Self-developed models, open-source models and deep modification scenarios should focus on determining service filing

If an enterprise develops or trains a large model itself, or deploys an open-source model that has not completed filing on its own servers and then provides services to the domestic public in its own name, this is usually no longer a simple application that "directly calls the capabilities of an already-filed model".

Ordinary prompt design, workflow orchestration, RAG knowledge base access and output format configuration for an already-filed model are usually not sufficient in themselves to change the model's filing status. However, if the enterprise carries out substantive training, fine-tuning or structural modification of the model such that it is no longer merely directly calling the capabilities of the originally filed model and forms an independent model service controlled by the enterprise itself, then further determination should be made according to the generative artificial intelligence service filing path.

At present, the publicly available rules have not yet given a fully quantified unified standard for "what degree of fine-tuning necessarily converts into service filing". Therefore, this question cannot be judged solely by whether the enterprise has done "debugging"; it should be comprehensively judged in light of factors such as whether the model weights have changed, whether the model capabilities have substantively changed, whether the enterprise independently controls the model and the responsibility for service output.

04. The difference between registration and filing is not just the amount of materials

Generative artificial intelligence service filing and application or function registration are not two names for the same procedure. Filing targets the generative artificial intelligence service itself, while registration targets applications or functions formed by directly calling the capabilities of already-filed models.

The Cyberspace Administration of China's announcement defines the registration target as generative artificial intelligence applications or functions that "directly call the capabilities of already-filed models through API interfaces or other means". From the regulatory logic, filing usually requires a more systematic explanation of model source, training data, service capabilities, security measures, content governance and risk control; registration, on the other hand, is built on the basis that the underlying model has already completed filing, and focuses on identifying who operates the new application or function, which model is called, what services are provided and how application-layer security responsibilities are implemented.

There are also differences between the two in terms of the competent authority and the level at which they are handled. Public announcements show that generative artificial intelligence applications or functions are registered by local cyberspace administration departments; enterprises providing relevant services usually need to submit applications, supplement materials and complete corresponding public disclosure in accordance with territorial regulatory requirements.

After completing filing or registration, launched applications should also publicly disclose in a prominent position or on the product details page the already-filed or registered generative artificial intelligence services used, and indicate the model name, filing number or launch number.

The distinction between filing and registration can be reduced to three questions: whether an already-filed model is directly called; whether the model is still merely applied without forming independently controlled model capabilities; and whether what is provided externally is the model service itself or a specific application or function built on a filed model.

05. The five misunderstandings most common among enterprises

Misunderstanding 1: Since the underlying large model has been filed, all products can go live directly

Completion of filing by the underlying model only shows that the model service has performed the corresponding procedures. When an enterprise embeds it into a new App, mini-program, website or agent and provides generative AI functions to the domestic public, it may still need to complete generative artificial intelligence application or function registration. The enterprise also needs to perform application-layer responsibilities for content security, personal information protection, labeling of generated content, user complaint handling and public disclosure of filing and registration information.

A filed model provides the precondition for application registration, but it cannot replace the application's own registration and ongoing compliance.

Misconception 2: As long as an API is used, only registration is required and filing can never be needed

An API is merely a technical means of invocation, not the sole legal criterion for distinguishing filing from registration. If an enterprise directly calls the capabilities of an already-filed model through an API and has not carried out substantive training or modification that results in the formation of independent model capabilities, it will generally enter the application or function registration path. However, if the model being called has not itself completed filing, or if the enterprise, although using an API as the technical interface, has conducted substantive training and in-depth modification of the model and formed independently controlled model capabilities, then the conclusion that only registration is required cannot be drawn merely from "using an API."

What really needs to be assessed is: whether the enterprise still constitutes a direct application of the capabilities of an already-filed model.

Misconception 3: If AI is used internally within an enterprise, large model filing must also be obtained

The Interim Measures for the Administration of Generative Artificial Intelligence Services exclude research and development and application activities that do not provide generative AI services to the domestic public. Therefore, where an enterprise provides AI tools only to its internal employees, this will generally not, on that basis alone, trigger generative AI service filing or application registration. However, internal use may still involve compliance risks relating to employee and customer personal information, enterprise trade secrets, third-party data processing, and cross-border data transfer.

The inapplicability of filing and registration cannot be understood to mean that internal enterprise AI is free from all constraints.

Misconception 4: After completing filing or registration, the product can operate according to the original plan

Filing and registration are procedural obligations, not an overall endorsement of the full legality of a product. After an AI product goes live, requirements concerning content security, personal information protection, complaints and reports, labeling of generated content, and handling of security incidents must still be implemented. Where there are major changes to product functions, model sources, or service models, it may also be necessary to handle changes or to reassess whether the original procedures still apply.

Filing, registration, and subsequent governance should be understood as a continuous process, not a one-time "compliance pass."

Misconception 5: After completing generative AI filing or registration, algorithm filing and security assessment are no longer required

Generative AI service filing, application registration, algorithm filing, and security assessment each address different regulatory issues and cannot substitute for one another. Providers of generative AI services with public opinion attributes or social mobilization capabilities shall also carry out security assessments in accordance with relevant national provisions and perform algorithm filing and change or cancellation procedures.⑥

After generative AI service filing or application registration is completed, it is still necessary to continue assessing whether security assessment and algorithm filing are triggered.

06. Enterprises can complete a preliminary assessment using four questions

Before an AI product or function goes live, an enterprise may make its assessment in the following order.

First, whether the enterprise uses generative AI technology to provide services to the domestic public in China. If it does not provide services to the domestic public, it will in principle not enter the generative AI service filing or application registration procedure, but it must still assess other data and algorithm compliance obligations.

Second, whether the enterprise directly calls the capabilities of an already-filed generative AI model. If it does not use an already-filed model but instead independently develops, trains, or self-deploys an unfiled model and provides services externally, it will in principle enter the generative AI service filing path.

Third, if the enterprise uses an already-filed model, it is necessary to assess whether the model has undergone substantive training or modification such that it no longer constitutes a direct call of the capabilities of the original already-filed model. If no independent model capabilities have been formed, generative artificial intelligence application or function registration will generally be carried out; if independently controlled model capabilities have been formed, further assessment under the service filing path is required.

Finally, for services that enter the filing or registration path, it should also be assessed whether they have public opinion attributes or social mobilization capabilities. If they have the corresponding attributes, a security assessment should also be carried out and algorithm filing performed.

This decision chart can only be used for general preliminary assessment. In practice, analysis must also be conducted in light of the product's service targets, control of model weights, training and fine-tuning methods, the scope of function availability, and the specific requirements of the local cyberspace administration department.

07. What preparations should enterprises make before handling filing or registration

Enterprises should not wait until product development is complete and launch is imminent to discuss filing or registration for the first time.⑦

At the product initiation or model integration stage, an enterprise first needs to sort out the business scenario and clarify the AI function's service targets, service region, user entry points, types of output content, and whether it is open to the unspecified public. For products that have both internal and external scenarios, assessments should be made separately and they should not be lumped together as "enterprise self-use" or "external services."

Second, the model source and filing status should be confirmed. Enterprises need to retain the model supplier, specific model name and version, filing information, API invocation method, and the scope of contractual authorization. A model cannot be deemed to have been filed merely on the basis of an oral promise by a supplier's salesperson.

Third, the modifications made by the enterprise to the model should be identified. There are clear differences between prompts, RAG, plug-ins, workflows, and ordinary parameter configuration on the one hand, and modification of model weights, continuous training, and structural fine-tuning on the other. Enterprises should require the technical team to produce an accurate explanation, so as to avoid operations that the technical department calls "fine-tuning" but that are legally merely application configuration, and also to avoid a situation where independent model capabilities have already been formed but are still packaged as simple API calls.

In terms of content security, enterprises need to prepare mechanisms for reviewing generated content, handling illegal and harmful information, managing user input, risk testing, complaints and reports, and human intervention.

In terms of data compliance, enterprises should sort out the personal information, sensitive personal information, trade secrets, and intellectual property-protected content involved in training data, knowledge base materials, user input, and model output, and clarify the sources, permissions, and boundaries of use.

Enterprises should also arrange for product, technical, legal, data compliance, and information security personnel to participate jointly in filing and registration work. The procedural obligations for generative AI are not simply filling out an application form, but a systematic explanation by the enterprise of its product functions, model sources, security capabilities, and operating mechanisms.

Conclusion: What enterprises need to manage is not an API, but an AI service facing users

After generative AI enters enterprise products, technical integration may take only a few days, but the change in regulatory identity often occurs in an instant.

When an enterprise uses third-party AI tools internally, it is usually a user of AI services; but when an enterprise packages model capabilities into its own products and directly provides AI functions to the public, it may become an application operator or service provider in the regulatory sense.

To determine whether filing or registration applies, one cannot look only at whether a foundational model was self-developed, nor only at whether it is called through an API. More importantly, one should look at: whether the enterprise provides services to the domestic public, whether it directly calls an already-filed model, and whether it has already formed independent model capabilities controlled by the enterprise itself.

As filing, registration, and special enforcement gradually become routine, the safest approach for enterprises is not to complete procedures after the product goes live, but to embed regulatory assessment into the AI product launch process at the stages of product design, model selection, and function development.

Article Appendix

① Cyberspace Administration of China: Notice on Launching the Special Action on "Qinglang · Cracking Down on Chaos in AI Applications", April 30, 2026.

② Cyberspace Administration of China: series of announcements on the Release of Information on Filed Generative Artificial Intelligence Services, 2024–2026. See the official website of the Cyberspace Administration of China: https://www.cac.gov.cn.

③ Cyberspace Administration of China: Provisions on the Administration of Algorithm Recommendations for Internet Information Services, issued December 31, 2021, effective March 1, 2022.

④ Cyberspace Administration of China: Provisions on the Administration of Deep Synthesis of Internet Information Services, issued November 25, 2022, effective January 10, 2023.

⑤ Cyberspace Administration of China and six other departments: Interim Measures for the Administration of Generative Artificial Intelligence Services, issued July 13, 2023, effective August 15, 2023.

⑥ The analysis in this article concerning "substantive training or modification forming independent model capabilities" is a practical judgment based on current rules and publicly available regulatory positions, and its specific application should be comprehensively assessed in light of product functions, model control methods, and local regulatory requirements.

⑦ Cyberspace Administration of China et al.: Measures for Labeling Artificial Intelligence-Generated Synthetic Content (effective September 1, 2025) and the supporting mandatory national standard (GB 45438-2025).

Questions and answers

Does an enterprise need to file or register if it only calls an already-filed large model API?

It depends on whether the enterprise provides generative AI services to the domestic public and whether it has conducted substantive training or modification. If it directly calls an already-filed model without forming independent model capabilities, application or function registration is generally required; if it has formed independently controlled model capabilities, service filing may be required.

What is the difference between generative AI service filing and application or function registration?

Filing targets the generative AI service itself and requires a systematic explanation of model source, training data, service capabilities, security measures, content governance and risk control. Registration targets applications or functions that directly call the capabilities of already-filed models, focusing on the operator, the model called, the services provided and application-layer security responsibilities.

Is filing or registration required for internal enterprise use of generative AI?

Generally no. The Interim Measures for the Administration of Generative Artificial Intelligence Services exclude research and development and application activities that do not provide services to the domestic public. Internal use by employees usually does not trigger filing or registration, but enterprises must still address personal information, trade secrets, third-party data processing and cross-border data transfer.

After completing generative AI filing or registration, are algorithm filing and security assessment still required?

Yes, if the service has public opinion attributes or social mobilization capabilities. Generative AI service filing, application registration, algorithm filing and security assessment address different regulatory issues and cannot substitute for one another. Providers must still carry out security assessments and perform algorithm filing procedures as required.

What four questions can an enterprise use to preliminarily assess whether AI filing or registration applies?

First, whether it provides generative AI services to the domestic public in China. Second, whether it directly calls an already-filed model. Third, if using an already-filed model, whether substantive training or modification has formed independent model capabilities. Finally, whether the service has public opinion attributes or social mobilization capabilities, triggering security assessment and algorithm filing.

Laws and cases cited

Laws

  • Cybersecurity Law
  • Data Security Law
  • Personal Information Protection Law

Regulations

  • Provisions on the Administration of Algorithm Recommendations for Internet Information Services
  • Provisions on the Administration of Deep Synthesis of Internet Information Services
  • Interim Measures for the Administration of Generative Artificial Intelligence Services
  • Measures for Labeling Artificial Intelligence-Generated Synthetic Content

Policy documents

  • Qinglang · Rectifying Chaos in AI Applications

Standards

  • GB 45438-2025