Analysis of the Comprehensive Impact of Tightening AI Data Regulation in China on Domestic AI Enterprises

Based on the latest policy trends and market conditions, I will systematically analyze the multi-dimensional impact of tightening AI data regulation on model training and commercial operations of domestic AI enterprises.

China’s AI regulation presents a governance concept of

. In July 2023, the Cyberspace Administration of China (CAC) jointly issued the Interim Measures for the Management of Generative Artificial Intelligence Services with seven departments, which is the [4]. On December 27, 2025, the CAC released the Interim Measures for the Management of Anthropomorphic Interactive Services of Artificial Intelligence (Draft for Comments), further tightening regulation on AI services with emotional interaction capabilities [1].

This series of regulatory policies forms a

with basic laws and regulations such as the Cyber Security Law, Data Security Law, and Personal Information Protection Law, constructing a compliance framework covering the entire life cycle of AI.

New regulations impose the following main requirements on AI enterprises in data processing activities:

• Dataset values compliance
  : Use datasets that conform to socialist core values for pre-training and optimization training
• Training data transparency
  : Enhance transparency and reliability explanations of training data sources
• Data diversity guarantee
  : Improve the diversity of training data and reduce algorithmic bias
• Synthetic data security assessment
  : Conduct security assessment on synthetic data
• Algorithm filing system
  : Conduct security assessment and algorithm filing for AI services with public opinion attributes or social mobilization capabilities
• Content identification obligation
  : Clearly identify AI-generated content

has become one of the biggest uncertainties faced by AI enterprises. Zhipu AI admitted in its prospectus:

“During business operations, we may obtain training data from third-party suppliers, public websites, public datasets or other publicly accessible sources. We cannot guarantee that these data are legal and compliant. If any of these entities fails or is deemed to have failed to obtain such data in a reasonable and legal manner, or fails to comply with applicable cyber security, data privacy and protection laws and regulations, it may have a significant adverse impact on our services and reputation.” [3]

This uncertainty is mainly reflected in:

1. Copyright infringement risk
   : A large amount of training data includes crawled Chinese web pages, social media content, news articles, etc., making it difficult to provide legal authorization or fair use basis [4]
2. Personal information protection
   : There are technical limitations in data desensitization and anonymization processing, which cannot guarantee full compliance with relevant laws and regulations [3]
3. Data quality control
   : Need to deduplicate, detoxify, and clean bias in data, increasing training complexity

According to national standards such as the Basic Requirements for the Security of Generative Artificial Intelligence Services and the Cybersecurity Technology—Security Specifications for Pre-training and Optimization Training Data of Generative Artificial Intelligence (GB/T 45652-2025) [4], AI enterprises must:

• Establish data traceability lists
  : Label each type of data with source channels, acquisition methods, legal basis, and data preprocessing processes
• Improve desensitization mechanisms
  : Effectively filter, desensitize, and anonymize data from third-party suppliers and publicly accessible sources before further processing
• Data quality assessment
  : Ensure the authenticity, accuracy, objectivity, and diversity of training data

These requirements lead to

, and enterprises like Zhipu AI admit that due to inherent limitations of data processing technology, it is difficult to ensure full compliance before processing [3].

Large model filing requires enterprises to clearly explain:

• Model type (base/fine-tuning/distillation)
• Training data composition (proportion, source, cleaning rules)
• Whether user data is used and how to desensitize it
• Inference deployment method (cloud/edge/mixed)
• Key parameters (training steps, token volume, fine-tuning methods) [4]

This requires enterprises to consider

in the early stage of model design, increasing the complexity of technical implementation.

According to McKinsey’s prediction, China will need

by 2030, but the domestic talent pool is expected to meet only about one-third of the demand, with a gap of about 4 million people [5]. This talent shortage, combined with increased technical complexity due to compliance requirements, further pushes up the labor costs of model training.

The

has become a necessary link for product launch. Industry practices show that 90% of teams encounter the following problems in preparing filing materials [4]:

• Security assessment report ‘useless’
  : Empty content, lack of methodology, insufficient test samples
• Algorithm description written as ‘product white paper’
  : Full of marketing words instead of technical facts
• Training data ‘unknown source’ or ‘copyright doubtful’
  : Cannot provide legal authorization basis
• User real-name and content review mechanism ‘paper talk’
  : Lack of implementable solutions
• Data storage and processing ‘closed loop unclear’
  : Mixed cloud deployment and use of overseas open source models lead to data出境 risks

These problems lead to frequent cases of

, and enterprises need an additional 2-3 weeks to revise them [4], significantly extending the product launch cycle.

AI enterprises need to invest a lot of resources to establish compliance systems:

• Establish AI governance committee
  : As the core decision-making body, coordinate compliance strategies and assume final responsibility
• Improve internal systems
  : Clarify the responsibility boundaries of various departments in AI projects, and establish a full-process approval mechanism from project initiation to deployment
• Technical compliance investment
  : Establish data annotation quality control, version traceability and change log systems to ensure that engineering decisions are closely tied to compliance requirements
• Regular review mechanism
  : Ensure that governance arrangements can be dynamically adjusted with technological evolution and regulatory updates [2]

These compliance costs constitute a heavy burden for AI startups that have not yet achieved profitability. Zhipu AI had revenue of 312 million yuan last year, but net loss reached 2.958 billion yuan; MiniMax’s annual loss was about 3.27 billion yuan [2].

Tightening regulation forces AI enterprises to re-examine their business models:

1. Shift from C-end to B-end
   : Zhipu AI focuses on the enterprise market, with 84.5% of its revenue last year coming from localized deployments in finance, government affairs and other fields, and remains the largest independent large model vendor in China with the “open source base + commercial customization” model [2]

2. Balance between global layout and local compliance
   : MiniMax chose global C-end breakthrough, with its AI community product Talkie covering more than 200 countries and attracting over 200 million users, but faces the trend of stricter content review in many countries [2]

3. Dilemma of ‘working for cloud vendors’
   : High computing power costs and talent competition make it difficult to get rid of dependence on cloud vendors in the short term [2]

AI enterprises face

:

• Intellectual property litigation
  : MiniMax is facing a joint lawsuit by the six major Hollywood studios, accusing its AI video tool Conch AI of infringing on film and television copyrights, with a maximum claim of 75 million US dollars [2]
• Data cross-border transmission compliance
  : Providing services to the EU requires compliance with strict data protection regulations such as GDPR [2]
• Algorithmic ethics review
  : Need to establish an algorithm impact assessment (AIA) mechanism [2]

These legal risks may not only lead to huge economic losses but also seriously affect enterprise reputation and market confidence.

Facing data compliance challenges, leading enterprises are reconstructing their data supply chains:

1. Build self-owned compliant datasets
   : Reduce dependence on third-party data and publicly crawled data
2. Establish authorized cooperation with copyright holders
   : Obtain legal authorization for training data
3. Invest in synthetic data technology
   : Develop synthetic data that meets security requirements
4. Establish data traceability system
   : Achieve full-process traceability from data collection to model training

Regulatory authorities in many countries have clearly required embedding compliance requirements into the entire product life cycle, which has become a

[2]. Enterprises should complete the following at the start of R&D:

• Legality review and risk identification
  : Conduct data protection impact assessment (DPIA) in accordance with Article 35 of GDPR
• Algorithm impact assessment
  : Refer to the requirements of the EU AI Act for high-risk systems
• Technical specification transformation
  : Transform assessment results into specific technical specifications

Facing homogeneous competition and compliance pressure, enterprises need to find differentiated paths:

• Deepen in vertical fields
  : Focus on industries with high data security requirements such as finance, medical care, and government affairs
• Layout edge AI
  : Develop end-to-end solutions to reduce data transmission and storage risks
• Build open source ecosystem
  : Establish technical moats and industry influence through open source

The cooling of the primary market forces enterprises to accelerate their breakthrough into the secondary market. Zhipu AI and MiniMax have successively submitted prospectuses to the Hong Kong Stock Exchange, launching the battle for the “first Chinese large model stock” [2]. However, analysts point out that fundamental issues such as

have not yet been resolved, and this IPO gamble has huge uncertainties [2].

Tightening AI regulation is accelerating industry shuffle:

• Head enterprises’ advantages become prominent
  : Enterprises with sufficient funds and technical strength are more capable of bearing compliance costs
• Survival pressure of startups increases
  : Enterprises with limited resources are difficult to meet complex compliance requirements
• Industry concentration increases
  : The market concentrates on enterprises with core AI algorithm capabilities and precision manufacturing experience

The Chinese government adopts a

regulatory approach [4], leaving space for technological innovation through methods such as “regulatory sandbox”. In August 2024, Beijing launched the AI+ initiative, aiming to apply AI to 90% of China’s economy by 2030 and eventually “reshape the paradigm of human production and life” [5]. This indicates that the regulatory goal is not to stifle innovation, but to guide technology to develop in a direction that conforms to national security and social values.

Although enterprises face compliance challenges in the short term, tightening regulation also brings positive impacts to the industry in the long run:

1. Improve overall industry quality
   : Clean up non-compliant small players and enhance the overall level of the industry
2. Enhance user trust
   : Compliance certification becomes a guarantee of product quality and safety
3. Promote benign competition
   : Benign competition based on technology and products replaces unfair competition in data acquisition
4. Drive technological innovation
   : Compliance pressure forces technological innovation, such as synthetic data and privacy computing technologies like federated learning

The tightening of AI data regulation in China has an

impact on domestic AI enterprises. In the short term, enterprises face challenges such as rising compliance costs, extended product launch cycles, and business model adjustments; in the long run, regulation will promote the industry to develop in a more standardized and healthy direction.

1. Regard compliance as core competitiveness
   : Layout compliance systems in advance and transform compliance capabilities into competitive advantages
2. Establish full-process data governance
   : Manage the entire life cycle from data collection, cleaning, training to deployment
3. Explore differentiated business models
   : Avoid homogeneous competition and focus on vertical fields and segmented markets
4. Strengthen intellectual property layout
   : Actively apply for patents and establish technical barriers
5. Pay attention to regulatory policy dynamics
   : Adjust strategies in a timely manner to adapt to changes in the regulatory environment

1. Attach importance to enterprise compliance capabilities
   : Take compliance as an important consideration in investment decisions
2. Rationally evaluate technological breakthroughs
   : Focus on real technical strength rather than marketing hype
3. Long-term value investment
   : AI is a long-term track, and patience is needed to wait for the maturity of technology and commercialization
4. Risk diversification
   : Avoid over-concentration of investment and reduce portfolio risks

The tightening of AI regulation is a “coming-of-age ceremony” necessary for the development of China’s AI industry. Enterprises that can find a balance between compliance and innovation will stand out in future competition.

[0] Jinling API Data

[1] Reuters - “China issues draft rules to regulate AI with human-like interaction” (https://www.reuters.com/world/asia-pacific/china-issues-drafts-rules-regulate-ai-with-human-like-interaction-2025-12-27/)

[2] Yahoo Finance Hong Kong - “Fight for the First Chinese Large Model Stock! Zhipu and MiniMax Successively Apply for Hong Kong IPO” (https://hk.finance.yahoo.com/news/搶當中國大模型第-股-智譜-minimax相繼申請赴港上市-專家直指-021002108.html)

[3] East Money - “National New Regulations: Platforms Shall Not Force Operators to Enable Automatic Price Following or Automatic Price Reduction丨Compliance Weekly” (https://finance.eastmoney.com/a/202512213597547448.html)

[4] Zhihu - “8 Pitfalls in Large Model Filing Materials, 90% of Teams Have Stepped On Them (With Avoidance Guide)” (https://zhuanlan.zhihu.com/p/1979124424784512323)

[5] TIME Magazine - “The Architects of AI Are TIME’s 2025 Person of the Year” (https://time.com/7339685/person-of-the-year-2025-ai-architects/)