Exploring Compliance Pathways For Intellectual Property In Large Model Training Data

Breakthrough advances in large model technologies have driven artificial intelligence from the “discriminative” era into the “generative” era. A key factor behind this leap in capability is the use of massive volumes of training data. However, existing legal frameworks governing intellectual property and data protection were not originally designed with large-scale machine learning and its extensive use of data in mind. The tension between existing legal frameworks and technological development has already given rise to numerous disputes and lawsuits in the large model industry. Between 2025 and 2026, landmark cases concerning alleged copyright infringement involving AI training data emerged in the United States, Germany, China and other jurisdictions, ranging from class actions brought by U.S. authors against six major AI companies, to a Munich court ruling that OpenAI infringed copyright by using song lyrics for model training without authorization, and the first copyright infringement case involving an AI large model adjudicated by the Shanghai Intellectual Property Court. Judicial practice is gradually delineating the legal boundaries governing the compliant use of training data.

Lawful Acquisition and Use of Third-Party Copyrighted Works

(I) Copyright Infringement Risks in Training Data

Large model training generally requires massive quantities of text, images, audio, video and other data to be copied onto training servers for preprocessing and model training. Such acts of “reproduction” may themselves fall within the scope of the reproduction right provided for under Article 10 of the Copyright Law of the People’s Republic of China. More importantly, where training data contains works protected by copyright and the developer has not obtained authorization from the copyright owner, the use of such works may constitute copyright infringement.

(II) Controversies over the Application of Fair Use

Whether the use of copyrighted works for large model training constitutes “fair use” is currently one of the most controversial legal issues in this field. Article 24 of China’s Copyright Law adopts a closed-list approach and specifies thirteen circumstances under which a work may be used without permission and without payment of remuneration. It is difficult, however, to bring the use of works for large model training within any of these enumerated circumstances through legal interpretation. Some scholars have argued that the use of pre-existing works for large model training “cannot be directly categorized as ‘fair use,’ nor can it be expansively interpreted as a form of ‘statutory license.’”

(III) Compliance Pathways

Given the uncertainty surrounding the application of the fair use doctrine, large model developers may establish a copyright compliance framework through the following approaches:

First, licensing and authorization. Developers may obtain lawful rights to use training data by entering into licensing agreements with copyright owners or copyright collective management organizations. This approach provides the greatest degree of legal certainty, but it also faces practical challenges, including high licensing costs, fragmented ownership of rights and low negotiation efficiency.

Second, prioritizing open-source and public-domain data. Developers may give priority to works that have entered the public domain and datasets made available under open-source licenses, thereby reducing copyright risks at the source.

Third, technical compliance measures. During data preprocessing, developers may screen training data for copyright-related risks and avoid using content expressly marked as prohibited for AI training. Technical measures may also be adopted during model training and output generation to reduce the risk of overfitting and minimize the “memorization” and reproduction of specific copyrighted works.

Legal Boundaries for the Processing of Personal Information

(I) Personal Information Risks in Training Data

A substantial proportion of the data used to train large models is sourced from the Internet and inevitably contains personal information, such as names, contact details, biometric information and behavioral records, and may even include sensitive personal information such as facial data. Once such information is incorporated into a training dataset, its use constitutes the “processing of personal information” regulated by the Personal Information Protection Law of the People’s Republic of China (“PIPL”).

(II) Legal and Regulatory Framework

Under Article 13 of the PIPL, personal information may be processed only where one of the following legal bases applies: (1) the individual’s consent has been obtained; (2) the processing is necessary for the conclusion or performance of a contract to which the individual is a party; (3) the processing is necessary for the performance of statutory duties or legal obligations; (4) the processing is necessary for responding to public health emergencies or protecting the life, health or property of natural persons under emergency circumstances; (5) the processing is carried out within a reasonable scope for purposes of news reporting, public opinion supervision or other activities conducted in the public interest; (6) personal information that has been disclosed by the individual or otherwise lawfully disclosed is processed within a reasonable scope in accordance with the PIPL; or (7) other circumstances prescribed by laws or administrative regulations.

For large model training, obtaining consent individually from every data subject represented in massive datasets is practically infeasible. At the same time, there are currently no clear standards as to how the “reasonable scope” for processing publicly disclosed personal information should be defined or whether large model training falls within such scope. Some scholars have suggested that, in relation to publicly disclosed personal information, the “purpose of disclosure” may be interpreted relatively broadly so as to permit its processing, whereas personal information that has not been publicly disclosed should continue to rely on individual consent as the legal basis for processing.

(III) Compliance Pathways

First, compliance review of data sources. Developers should review and filter the sources of training data and refrain from using data obtained from unlawful sources. At the data collection stage, developers should assess whether the data contains personal information, particularly sensitive personal information.

Second, anonymization and de-identification. During data preprocessing, personal information should be anonymized or de-identified so that specific natural persons cannot be identified and the information cannot be restored, thereby removing such data from the regulatory scope of the PIPL.

Third, fulfillment of notice-and-consent obligations. Where the processing of personal information is necessary, developers should fulfill the statutory notice requirements and obtain consent in accordance with law, while ensuring compliance with the principles of lawfulness, legitimacy, necessity and transparency.

Protection and Compliance Relating to Patent-Related Technical Information and Trade Secrets

(I) Dual Considerations: Infringement Risks and Protection of Proprietary Assets

Issues relating to patents and trade secrets in large model training data have two dimensions. On the one hand, developers may face infringement risks where training data contains third-party patented technical information or trade secrets. On the other hand, a developer’s own core technologies for large models, such as model architectures, training-data selection rules and parameter optimization methods, may also require protection through patent or trade secret regimes.

(II) Infringement Risks Relating to Third-Party Technical Information

Where training data contains third-party patented technical solutions, unpublished research results, proprietary corporate technical documents or similar materials, the use of such data may constitute patent infringement or the misappropriation of trade secrets. Article 9 of the Anti-Unfair Competition Law prohibits the acquisition of a rights holder’s trade secrets through theft, bribery, fraud, coercion, electronic intrusion or other improper means.

(III) Protection of Proprietary Core Technologies

From the perspective of protecting an enterprise’s own assets, core assets relating to large models, including model architecture design, training-data selection rules, hyperparameter configuration logic and feature-extraction algorithms, may be protected as trade secrets. Unlike copyright and patent protection, which are subject to statutory terms of protection, trade secret protection may continue indefinitely, provided that the relevant information remains confidential.

(IV) Compliance Pathways

First, screening data sources for patent-related information and trade secrets. During data collection and preprocessing, developers should screen data for potentially patented technical information, unpublished technical documents and other sensitive materials, so as to avoid the improper acquisition of third-party technical secrets.

Second, establishing a trade secret management system. Reasonable confidentiality measures should be implemented for proprietary core algorithms, training data, parameter configurations and other sensitive information, including the execution of confidentiality agreements, access controls and information classification systems.

Third, strategic selection between patent protection and trade secret protection. For technical solutions that can be readily reverse-engineered, patent protection should be given priority.

Accordingly, intellectual property compliance for large model training data cannot be achieved through a single legal mechanism or a single compliance pathway. Instead, an integrated framework featuring “classified governance and stage-based controls” should be established.

(I) Classified Governance: Differentiating Data Types and Legal Attributes

Different categories of training data are subject to different legal regimes and should therefore be addressed through differentiated compliance strategies:

• Copyrighted works: Priority should be given to licensed or open-source data. Given the current uncertainty surrounding the application of fair use, the legal risks associated with unauthorized use should be carefully assessed.

• Personal information: Anonymization should be the preferred approach. Where processing is necessary, statutory notice-and-consent obligations and impact assessment requirements should be strictly observed.

• Patent-related technical information and trade secrets: Mechanisms should be established to screen data sources and prevent the improper acquisition of third-party technical secrets, while enterprises should also strengthen their internal trade secret management systems.

(II) Stage-Based Controls: Covering the Entire Data Lifecycle

Compliance relating to large model training data should extend throughout the entire lifecycle, including data collection, preprocessing, training, deployment and output generation:

• Collection stage: Review the lawfulness of data sources and assess risks relating to copyright, personal information and technical information.

• Preprocessing stage: Conduct copyright screening, anonymize personal information and desensitize technical information.

• Training stage: Ensure that the training process complies with applicable intellectual property laws and pay close attention to the conditions governing the application of fair use.

• Deployment and output stage: Prevent model outputs from reproducing third-party copyrighted works, personal information or technical solutions.

Intellectual property compliance for large model training data is an institutional threshold that must be addressed to ensure the healthy development of the artificial intelligence industry. Across the three dimensions of third-party copyright, personal information, and patent-related technical information, existing legal frameworks remain insufficient to varying degrees. Addressing these challenges requires legislators to advance institutional reform in response to technological developments, developers to establish prudent compliance systems within the existing legal framework, and judicial authorities to progressively clarify legal boundaries through individual cases. Only by achieving a dynamic balance between incentives for innovation and the protection of rights and interests can the high-quality development of the artificial intelligence industry be aligned with the rule of law.