Introduction

The rapid development of generative artificial intelligence has made "whether public internet data can be used directly for AI training" a frequent question in corporate R&D, commercialization and compliance review. The intuition of many companies is: since the relevant data has already been published publicly on the internet, and is technically accessible, crawlable, cleanable and trainable, it should seemingly also be permitted to use legally. But this judgment is not reliable. Public availability does not of course equal freedom from rights burdens, and being able to access does not of course mean it can be used for commercial model training.

From the EU Artificial Intelligence Act (AI Act) to China's Interim Measures for the Administration of Generative Artificial Intelligence Services, the regulatory logic is becoming clearer: the compliance focus of AI training data has extended from a sole focus on model output to the entire process, including data sources, collection methods, training processes, rights restrictions, personal information protection, content labeling and accountability records.

The author of this article, Zhu Kai, senior partner at the Shanghai office of Jincheng Tongda & Neal Law Firm, believes that the core issue of AI training data compliance is not "whether the data has already been made public", but whether the enterprise can prove that its acquisition, processing and use of such data have a lawful source, legitimate purpose, reasonable boundaries and traceable governance measures. For enterprises, the real risk is often not "using internet data", but that when regulators, customers, investors or rights holders ask, the enterprise cannot explain "where these data came from, on what basis they can be used, where they have been used, how they have been processed, and whether they can be deleted or discontinued".

Public internet data usually includes news reports, social media content, forum comments, public web pages, images, audio and video, paper abstracts, product reviews, recruitment information, corporate website information, etc. Their common feature is that they can be accessed by the public, but their legal attributes are not the same. Some data may constitute personal information, some content may be protected by copyright, neighboring rights, database rights or platform rules, and some may also involve trade secrets, sensitive information, important data or national security risks.

In the dimension of personal information protection, Article 13 of China's Personal Information Protection Law provides that the processing of personal information shall have a statutory ground; although Article 27 permits processing, within a reasonable scope, of personal information that an individual has made public by himself or that has otherwise been lawfully made public, it also provides that processing is not permitted where the individual has explicitly refused; and where the processing activity has a major impact on the rights and interests of the individual, the individual's consent shall still be obtained in accordance with law.[4] This means that "public personal information" is not an exempt zone for personal information processing.

In AI training scenarios, the judgment of "reasonable scope" is particularly important. Enterprises cannot look only at whether the information is public; they should also review whether the purpose of use deviates from the original public scenario and whether the processing method exceeds the individual's reasonable expectations. For example, a job seeker makes his resume public on a recruitment platform for job matching; if the relevant information is bulk-crawled and then used to train sentiment analysis, consumption capacity assessment or employee stability prediction models, it may exceed the original public purpose and reasonable expectations. Another example: a user makes his avatar, nickname and comments public on a social platform, which does not of course mean that he agrees to the relevant data being included on a long-term basis in commercial large-model training. The more specific the public scenario and the more generalized the subsequent training use, the more prudent the assessment of its legal basis needs to be.

In the dimension of intellectual property, articles, images, music, videos, code, photographic works, etc. on public web pages are often still protected by copyright or other rights and interests. Public dissemination does not mean the rights holder has waived its rights, nor does it mean a third party may bulk-copy, download, cache and use them for commercial training. The legal risk of AI training does not arise only at the model output stage. Large-scale reproduction, cleaning, annotation, vectorization and dataset fixation at the front end of training may equally give rise to rights disputes.

Therefore, "whether public internet data can be used for AI training" cannot be answered with a simple yes or no. A more accurate conclusion should be: public data can become one of the sources of training data, but it must undergo multiple reviews covering lawful source, personal information, intellectual property, platform rules and data security.

02. The EU AI Act approach: not a blanket prohibition, but strengthened transparency and accountability

The EU AI Act does not take the position of "blanket prohibition of using public internet data to train AI". Its institutional design is closer to permitting technological development while imposing auditable and accountable obligations on model providers through mechanisms such as risk classification, transparency, technical documentation, copyright compliance policies and summaries of training content.

For providers of general-purpose AI models, Article 53 of the AI Act requires them to draw up and keep up to date technical documentation, including information on the training and testing process; at the same time, they shall put in place a policy to comply with EU copyright law and publish, in accordance with the template provided by the AI Office, a sufficiently detailed summary of the training content.[1] This shows that EU regulation does not simply ask "whether internet data was used", but requires model providers to explain "what types of data were used, how they were obtained, how rights restrictions were respected, and how training data governance records were formed".

This institutional arrangement is closely related to the text and data mining rules in the EU Digital Single Market Copyright Directive (DSM Copyright Directive). Article 3 of that directive is mainly aimed at text and data mining for scientific research purposes by research organizations and cultural heritage institutions; Article 4 provides an exception for broader text and data mining, but allows rights holders to reserve their rights in an appropriate manner.[3] For commercial entities, this means they cannot simply ignore rights reservation arrangements made by rights holders on the ground that "the data is public".

At the same time, Article 2 of the AI Act sets out its scope of application. For providers placing AI systems or general-purpose AI models on the EU market or putting them into service, even if their establishment is located outside the EU, they may be affected by the AI Act.[2] Therefore, Chinese enterprises intending to enter the EU market or provide models, systems or API services to EU users should note that the AI Act has a certain extraterritorial effect in its scope of application.

Based on the project experience of Zhu Kai's team, the biggest lesson of the EU AI Act for Chinese enterprises is that training data compliance is shifting from "result liability" to "process liability". Enterprises must not only avoid model output of illegal and infringing content, but also be able to prove that training data sources, cleaning rules, identification of rights reservations, dataset version management and deletion mechanisms have a compliance basis.

03. China's generative AI regulation: lawful source is the entry point, multiple laws are the bottom line

China's regulation of generative artificial intelligence is not a single-department-law path, but is jointly constituted by cybersecurity, data security, personal information protection, algorithm governance, deep synthesis, generative AI service management and labeling rules for generated and synthesized content.[7]

As for training data, the most direct basis is Article 7 of the Interim Measures for the Administration of Generative Artificial Intelligence Services. That article explicitly requires that providers carrying out training data processing activities such as pre-training and optimization training shall use data and foundation models with lawful sources; where intellectual property is involved, they shall not infringe the intellectual property rights lawfully enjoyed by others; where personal information is involved, they shall obtain the individual's consent or comply with other circumstances provided by laws and administrative regulations; and they shall take effective measures to improve the quality of training data and enhance the authenticity, accuracy, objectivity and diversity of training data.[5]

This article establishes the basic framework for China's generative AI training data compliance: lawful source is the entry point, intellectual property and personal information are the bottom line, data quality is a governance requirement, and basic laws such as the Cybersecurity Law, the Data Security Law and the Personal Information Protection Law provide the external boundaries.

Article 3 of the Data Security Law defines data processing as including activities such as collection, storage, use, processing, transmission, provision and disclosure of data; Article 27 requires that data processing activities be carried out with a sound full-process data security management system.[6] For AI training, crawling is only the starting point of the data processing chain. Subsequent cleaning, annotation, vectorization, training, evaluation, retraining, model release, API calls and log return may all constitute data processing activities. Enterprises cannot review only where the data comes from; they must also review how the data flows, who can access it, how it is graded and classified, whether it can be deleted, and whether it can be traced.

In addition, the Measures for Labeling AI-Generated Synthetic Content issued in 2025 and the national standard GB45438-2025 Cybersecurity Technology — Labeling Method for Content Generated and Synthesized by Artificial Intelligence further strengthen the requirements for explicit and implicit labeling of generated and synthesized content.[8][9] Although the labeling rules do not directly resolve the issue of training data sources, their regulatory logic is consistent: generative AI services cannot be regarded as ordinary technical tools, but should establish a full-process governance mechanism covering data sources, model development, service provision, content generation, dissemination and distribution, and user rights protection.

04. Comparison of the compliance logic for AI training data in China and Europe

To facilitate understanding by corporate legal counsel and management, the regulatory logic of China and Europe can be compared as follows:

It can be seen that the EU emphasizes more "explainable process liability" and "transparency obligations", while China emphasizes more the "bottom line of lawful source" and "coordinated governance under multiple laws". But the common trend of the two is consistent: public internet data cannot be used as an unconditional training resource, and enterprises must establish a data governance system that can be reviewed, verified and held accountable.

05. Practical risks for enterprises: insufficient compliance proof capability is the core risk

In corporate projects, the most common risk of training on public internet data is not subjective malicious infringement by the enterprise, but the inability to prove its own compliance. Many enterprises' datasets come from crawlers in the early R&D stage, open-source datasets, third-party suppliers, accumulated historical projects or personal collection by engineers. By the time the product is commercialized, financed, listed, taken overseas, or subject to procurement review by a major customer, it is discovered that the training data lacks source records, authorization chains, collection times, URL traces, identification of rights restrictions, personal information filtering records and deletion mechanisms.

The first type of risk is personal information risk.

Large-model training may cross-link public personal information with other data sources, forming new profiles, new inferences or new uses beyond the original public scenario. If the training data contains sensitive personal information such as facial images, medical health, financial accounts, whereabouts trajectories and information on minors, compliance requirements will be significantly higher.

The second type of risk is intellectual property risk.

For text, image, music, video and code models, training data and works are highly intertwined with rights. If an enterprise cannot explain the data source, the scope of authorization, and whether content that prohibits scraping or prohibits training has been excluded, it may, after commercialization, face rights holders' enforcement actions, platform claims, liability under customer contracts, or regulatory attention. This is especially so for enterprises targeting overseas markets: the EU copyright reservation mechanism, the copyright disputes in the United States over generative AI training data, and the differences between jurisdictions in the regimes for fair use and the text and data mining exception may all complicate compliance assessments. [10][11][12]

The third category of risk is the risk of suppliers and open-source datasets.

When an enterprise purchases third-party datasets or uses open-source datasets, it cannot discharge its compliance obligations merely on the strength of a single contractual sentence that "the supplier guarantees the data is lawful." The enterprise still needs to review the data source, authorization documents, license agreements, terms prohibiting commercial use, sublicensing restrictions, the basis for processing personal information, and data deletion arrangements.

The fourth category of risk is the risk of platform rules and unfair competition.

Although some websites are publicly accessible, their robots protocols, user agreements or developer rules may restrict bulk scraping, commercial use or AI training purposes. Violating platform rules does not necessarily amount to illegality, but in specific scenarios it may give rise to risks such as breach of contract, unfair competition, circumvention of technical measures, and disputes over commercial data rights and interests. Enterprises should in particular avoid high-frequency scraping, bypassing access restrictions, scraping content behind a login, or scraping data that is clearly not open to the public.

06. Corporate compliance recommendations: from "can it be used" to "how can it be proven that it can be used"

Lawyer Zhu Kai believes that enterprises should not simply form a polarized judgment of "public data cannot be used" or "all public data can be used." A more operational compliance boundary is: public internet data may be used for AI training provided that lawful source, reasonable scope, respect for rights, necessary filtering, security governance and auditable records are satisfied; once personal information, protected works, data in sensitive fields, minors' data, content restricted by platforms, or cross-border business are involved, the level of compliance review should be raised.

Drawing on the experience of Lawyer Zhu Kai's team in AI-related project transactions, IPO data compliance reviews and corporate data governance projects, enterprises can implement AI training data compliance as a "four-step data compliance" workflow.

(1) Conduct gatekeeping review before ingestion (Gatekeeping). Enterprises should establish a training data gatekeeping mechanism, setting separate review standards for their own business data, public internet data, authorized data, open-source datasets, government open data, third-party purchased data and user input data. The review covers the data source, scope of authorization, license agreements, commercial use restrictions, rights holders' opt-out, platform rules, personal information attributes, sensitive information attributes and cross-border use restrictions. Data that has not been reviewed should not directly enter the training data pool.

(2) Conduct dual filtering during cleaning (Processing). At the personal information level, de-identification, anonymization or minimization should be carried out before training to the greatest extent possible, and stricter exclusion rules should be established for sensitive personal information, minors' information, contact details, facial images, etc. At the intellectual property level, a removal mechanism should be established for clearly protected works, paid content, member-only content, and content that declares a prohibition on scraping or a prohibition on AI training. For content with high commercial value or concentrated ownership, data should preferably be obtained through authorization, cooperation or procurement.

(3) Establish a dynamic compliance ledger during training (Auditing). The lawful source ledger for training data should not be merely a one-off registration form, but a dynamic management tool running through the entire data lifecycle. The ledger should at least record the data source, collection method, collection time, collecting entity, authorization documents or license agreements, whether personal information is involved, whether intellectual property is involved, whether there are restrictions prohibiting training or prohibiting commercial use, cleaning and filtering measures, dataset versions, corresponding models, and deletion or suspension records. Its core function is to help enterprises prove, when regulatory inquiries, customer due diligence, investment reviews or rights disputes arise, that "the source is lawful, the processing is well-founded, and the use is controllable."

(4) Implement content governance at the output and dissemination stage (Output). At the stages of model release, API services, product launch and content distribution, enterprises should simultaneously implement the obligations of explicit labeling and implicit labeling of generated synthetic content, and establish mechanisms for complaints and reports, content handling, log retention and response to user rights. Training data compliance is not a single-point task of the R&D department, but a systematic project running through model R&D, product operations, customer delivery and dispute response. Enterprises should also embed training data compliance into contract governance. When procuring data, models or algorithm services externally, they should specify data source guarantees, scope of authorization, personal information processing responsibilities, intellectual property infringement indemnities, regulatory cooperation, data deletion, audit rights and liability for breach. Internally, they should establish R&D admission procedures and prohibit R&D personnel from arbitrarily scraping, downloading or mixing data of unknown origin in their own names.

Conclusion

Public internet data is an important resource for AI development, but it is not a "resource without an owner" in the legal sense. The EU AI Act, through the obligations for general-purpose AI models, the copyright compliance policy and the training content summary system, pushes enterprises from "black-box training" toward "explainable, provable and auditable" training data governance. China's Interim Measures for the Administration of Generative Artificial Intelligence Services, with lawful source, intellectual property, personal information and data quality at their core, establish the basic compliance bottom line for generative AI training data.

For enterprises, the key to AI training data compliance lies not in finding an abstract answer, but in establishing an evidence system that can withstand scrutiny by regulators, customers, investors and dispute proceedings. Whether public data can be used for training ultimately depends on whether the enterprise can prove that: the data source is lawful, the purpose of use is legitimate, the scope of processing is reasonable, rights restrictions are respected, individual rights and interests are protected, and security measures remain effective.

In the future, the competitiveness of AI enterprises will come not only from model capabilities, but also from training data governance capabilities. Whoever can establish a credible, transparent and auditable data compliance system earlier will be more likely to obtain a stable legal foundation and market trust in the commercialization and international competition of generative AI.

Questions and answers

Can public internet data be used directly for AI training under China's Generative AI Measures?

No, not directly. Article 7 of the Interim Measures requires providers to use data and foundation models with lawful sources, respect intellectual property, obtain personal information consent or comply with other legal circumstances, and improve data quality. Public availability alone does not satisfy these requirements.

What does the EU AI Act require for general-purpose AI model training data?

Article 53 requires providers to draw up technical documentation on training and testing, put in place a copyright compliance policy, and publish a sufficiently detailed summary of training content. The EU does not ban public data use but imposes transparency and accountability obligations.

Does public personal information fall outside the scope of China's Personal Information Protection Law?

No. Article 27 permits processing public personal information within a reasonable scope but prohibits processing if the individual explicitly refuses, and requires consent where processing has a major impact on individual rights. Public personal information is not an exempt zone.

What are the main practical risks for enterprises using public internet data for AI training?

The core risk is inability to prove compliance. Enterprises may face personal information risks from cross-linking, intellectual property risks from protected works, supplier and open-source dataset risks, and platform rule or unfair competition risks from scraping restrictions.

What compliance steps should enterprises take for AI training data?

Enterprises should implement a four-step workflow: gatekeeping review before ingestion, dual filtering during cleaning for personal information and intellectual property, a dynamic compliance ledger during training, and content governance at output including labeling and complaint mechanisms.

Laws and cases cited

Laws

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

Regulations

  • EU Artificial Intelligence Act (AI Act)
  • Interim Measures for the Administration of Generative Artificial Intelligence Services
  • EU Digital Single Market Copyright Directive (DSM Copyright Directive)
  • Measures for Labeling AI-Generated Synthetic Content

Standards

  • GB45438-2025 Cybersecurity Technology — Labeling Method for Content Generated and Synthesized by Artificial Intelligence

References

  1. Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act), Article 53. This article provides that providers of general-purpose AI models should draw up and keep up to date technical documentation, adopt a policy to comply with EU copyright law, and publish a summary of the training content.
  2. Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 2. This article sets out the scope of application of the AI Act, including situations such as placing AI systems on the EU market or putting them into service, and general-purpose AI models. The official text is available on EUR-Lex.
  3. Directive (EU) 2019/790 of the European Parliament and of the Council of 17 April 2019 on copyright and related rights in the Digital Single Market, Articles 3 and 4. This directive provides for the text and data mining exception and the rights holders' reservation mechanism. The official text is available on EUR-Lex.
  4. Articles 13, 27, 28 and 29 of the Personal Information Protection Law of the People's Republic of China. Article 13 provides the legal bases for processing personal information; Article 27 provides that personal information processors may process personal information that an individual has made public on their own or that has otherwise been lawfully made public within a reasonable scope, but where the individual expressly refuses or the processing activity has a major impact on the individual's rights and interests, corresponding restrictions shall apply.
  5. Article 7 of the Interim Measures for the Administration of Generative Artificial Intelligence Services, Order No. 15 of the Cyberspace Administration of China and six other departments, promulgated on 13 July 2023, effective from 15 August 2023. Article 7 provides that providers carrying out training data processing activities such as pre-training and optimization training shall use data and foundation models with lawful sources, shall not infringe intellectual property, shall obtain individual consent or meet statutory circumstances where personal information is involved, and shall improve the quality of training data.
  6. Articles 3 and 27 of the Data Security Law of the People's Republic of China. Article 3 provides that data processing includes activities such as collection, storage, use, processing, transmission, provision and disclosure of data; Article 27 provides that data processing activities shall be carried out with a sound full-process data security management system.
  7. The Provisions on the Administration of Algorithm Recommendations for Internet Information Services, the Provisions on the Administration of Deep Synthesis of Internet Information Services and the Interim Measures for the Administration of Generative Artificial Intelligence Services together constitute an important regulatory foundation for China's generative artificial intelligence services. Among them, the Provisions on the Administration of Deep Synthesis of Internet Information Services were issued by the Cyberspace Administration of China, the Ministry of Industry and Information Technology and the Ministry of Public Security, and took effect on 10 January 2023.
  8. The Measures for Labeling Artificial Intelligence-Generated Synthetic Content, jointly issued by the Cyberspace Administration of China, the Ministry of Industry and Information Technology, the Ministry of Public Security and the National Radio and Television Administration, promulgated on 14 March 2025, effective from 1 September 2025.
  9. National Standard GB 45438-2025 Cybersecurity Technology—Labeling Method for Artificial Intelligence-Generated Synthetic Content, issued on 28 February 2025, implemented on 1 September 2025, with the competent department and the department in charge both being the Office of the Central Cyberspace Affairs Commission, and the issuing units being the State Administration for Market Regulation and the Standardization Administration of China.
  10. U.S. Copyright Office, Copyright and Artificial Intelligence, Part 3: Generative AI Training, Pre-publication Version, May 2025. This report discusses the U.S. copyright law issues involved in using copyrighted works to train generative AI systems, and may serve as an official reference for understanding the U.S. copyright disputes over AI training.
  11. Reuters, NY Times sues OpenAI, Microsoft for infringing copyrighted work, 27 December 2023. This report concerns The New York Times suing OpenAI and Microsoft, alleging that they used news works without permission to train chatbots and raising other disputes.
  12. Reuters, Getty Images lawsuit says Stability AI misused photos to train AI, 6 February 2023. This report concerns a copyright lawsuit brought by Getty Images against Stability AI, alleging that it copied and used Getty images without permission to train Stable Diffusion.