Panorama of Large Language Model Training Data Compliance: Three Major Risks of Copyright, Personal Information and Trade Secrets and Their Practical Implementation Paths

The leap in capabilities of generative AI is essentially driven by massive volumes of training data. However, current intellectual property and data protection rules are not designed for large-scale machine learning scenarios. The gap between institutions and technologies has made training data compliance a Sword of Damocles hanging over all large model developers.

Between 2025 and 2026, landmark cases involving copyright infringement of AI training data have emerged successively in the United States, Germany, China and other regions. From the collective lawsuit filed by American writers against six major AI manufacturers, and the ruling of the Munich Court in Germany that OpenAI's unauthorized use of lyrics for training constitutes infringement, to the first domestic AI large model copyright infringement case concluded by the Shanghai Intellectual Property Court, global judicial practice is gradually outlining the legal boundaries of training data.

This article analyzes the infringement risks and compliance paths of training data from three core dimensions: third-party copyright, personal information, patents and trade secrets, and finally constructs a systematic compliance framework.

Third-Party Copyright: Contested Boundaries Between Infringement Risks and Fair Use

(1) The training act itself may trigger the risk of reproduction right infringement

The basic link of large model training is to copy massive amounts of text, images, audio, video and other data to the training server for preprocessing and model iteration. This "reproduction" act itself falls within the scope of control of the reproduction right stipulated in Article 10 of the Copyright Law. If the training data contains works protected by copyright, and the developer has not obtained permission from the right holder, there is a direct risk of infringement.

(2) Fair Use: The Most Controversial Legal Gap

Whether the use of copyrighted works in large model training can constitute "fair use" is the core of current industry disputes. Article 24 of China's Copyright Law adopts the legislative model of closed enumeration, which clearly stipulates 13 situations where works can be used without permission and without paying remuneration. However, the technical attributes of large model training can hardly be directly classified into any of these situations. Some views in the industry hold that this act "can neither be directly classified as 'fair use', nor can it be extendedly interpreted as 'statutory license'", and the legal characterization has a high degree of uncertainty.

(3) Three Types of Compliance Path Options

Facing the uncertainty of the fair use system, large model developers can build a copyright compliance system from three paths:

First, the authorization and licensing model. Sign licensing agreements with copyright owners and copyright collective management organizations to obtain the legal right to use training data. This is the most legally reliable path, but in reality it faces obstacles such as scattered right holders, high negotiation costs, and low efficiency.

Second, prioritize the use of open source and public domain data. Prioritize the selection of works that have entered the public domain and public datasets with open source license agreements to avoid copyright risks from the source.

Third, technical compliance measures. Conduct copyright screening in the data preprocessing stage to avoid content clearly marked as "prohibited for AI training"; adopt technical means in the training and output stages to reduce the risk of overfitting, and reduce the model's "memory" and reproduction of specific copyrighted works.

Personal Information: Legal Boundaries of Data Processing

(1) Training data inherently carries personal information risks

Most of the training data of large models are collected from the Internet, which inevitably contains personal information such as names, contact information, biometric information, behavior tracks, and even sensitive personal information such as human faces and health data. Once this information enters the training dataset, it constitutes a "personal information processing act" regulated by personal information protection laws, triggering a full set of compliance obligations.

(2) Practical Dilemma of Law Application

According to Article 13 of the Personal Information Protection Law, the processing of personal information must have one of the following legal bases: (1) Obtain the consent of the individual; (2) Necessary for concluding or performing a contract; (3) Necessary for performing statutory duties or statutory obligations; (4) Necessary for responding to public health emergencies or emergency situations; (5) Processing personal information within a reasonable range for news reporting, public opinion supervision and other acts for the public interest; (6) Processing personal information that is publicly disclosed by the individual or other legally publicly disclosed personal information within a reasonable range in accordance with the provisions of this Law; (7) Other circumstances stipulated by laws and administrative regulations.

But for large model training:

It is almost not operationally feasible to obtain the consent of each information subject in the massive data one by one;

For publicly disclosed personal information, there are currently no clear judicial or law enforcement standards on how to define the "reasonable range" and whether large model training falls within this range.

Some scholars suggest that for publicly disclosed personal information, it can be included in the scope of processable information through a lenient interpretation of the "purpose of disclosure"; for undisclosed personal information, personal consent is still required as the source of legality.

(3) Core Compliance Paths

First, compliance review of data sources. Carry out source verification and risk assessment at the stage of data collection, filter data from illegal sources, and focus on identifying and managing sensitive personal information.

Second, anonymization and desensitization processing. Anonymize or desensitize personal information in the preprocessing stage to meet the standard of being unable to identify a specific natural person and cannot be restored, so that it no longer falls into the regulatory scope of personal information.

Third, fulfill the "notification-consent" obligation in accordance with the law. For scenarios where personal information processing is indeed necessary, strictly follow the principles of legality, propriety, minimum necessity, openness and transparency, fulfill the notification obligation and obtain corresponding consent.

Patents and Trade Secrets: Dual-Oriented Risks and Protection

The issue of patents and trade secrets of large model training data presents a typical "dual orientation": it is necessary to prevent infringement caused by data containing other people's technical information, and also do a good job in intellectual property protection of its own core technologies.

(1) Infringement Risk: Improper Acquisition of Other People's Technical Information

If the training data contains other people's patented technical solutions, unpublished scientific research results, enterprise-specific technical documents and other content, it may constitute infringement of other people's patent rights, or obtain trade secrets by improper means, violating the relevant provisions of Article 9 of the Anti-Unfair Competition Law.

(2) Self-protection: Confidentiality Layout of Core Technologies

From the perspective of enterprise self-protection, the core assets of large models, including model architecture design, training data screening rules, hyperparameter configuration logic, feature extraction algorithms, etc., can all be protected through the trade secret system.

Different from the fixed protection period of patents and copyrights, trade secrets have no statutory protection period. As long as the information is always kept confidential, it can continue to be protected.

(4) Compliance and Protection Paths

First, data source screening. At the stage of data collection and preprocessing, screen for patented technical information and unpublished technical documents that may be contained in the data to avoid improperly obtaining other people's technical secrets.

Second, establish a confidentiality management system. Adopt hierarchical confidentiality measures for its own core algorithms, training data, parameter configuration, etc., including signing confidentiality agreements, setting access permissions, and establishing an information hierarchical management system.

Third, strategic choice of patent layout and trade secrets. For technical solutions that are easy to be reverse engineered, prioritize patent protection; for internal technical logic that is difficult to be cracked by the outside world, trade secret protection can be prioritized.

Systematic Compliance: Classified Governance + Phased Control

The intellectual property compliance of large model training data cannot be solved by a single legal tool or a single path, and it is necessary to build a comprehensive framework of "classified governance and phased control".

(1) Classified Governance: Matching Compliance Strategies According to Data Types

Different types of training data correspond to different legal regulations, and differentiated compliance strategies should be adopted:

Copyrighted Works: Prioritize the use of authorized licenses or open source/public domain data, and carefully assess the legal risks of unauthorized use under the background that the fair use rules are not yet clear.

Personal Information: Take anonymization and desensitization as the preferred solution, and strictly perform statutory procedures and impact assessments when processing is indeed necessary.

Patented Technical Information and Trade Secrets: Establish a normalized data source screening mechanism, and at the same time improve its own trade secret protection system.

(2) Phased Control: Cover the Full Life Cycle of Data

Data compliance should run through the whole process of data collection, preprocessing, training, deployment and output:

Collection Stage: Review the legality of data sources, and simultaneously assess three types of risks: copyright, personal information and technical information.

Preprocessing Stage: Implement copyright screening, personal information anonymization, technical information desensitization and other processing.

Training Stage: Ensure that the training process complies with intellectual property rules, and continuously pay attention to the judicial dynamics of systems such as fair use.

Deployment and Output Stage: Prevent the model output from reproducing other people's copyrighted works, personal information or confidential technical solutions.

Conclusion

The intellectual property compliance of large model training data is an institutional threshold that must be crossed for the high-quality development of the artificial intelligence industry. At present, there are varying degrees of supply gaps in existing laws in the three dimensions of third-party copyright, personal information, patents and trade secrets. To meet this challenge, it is not only necessary to promote institutional improvement at the legislative level in combination with technological development, but also for enterprises to build a prudent compliance system within the current legal framework. At the same time, judicial practice is gradually clarifying legal boundaries in individual cases. Only by finding a dynamic balance between innovation incentives and rights protection can the organic unity of the development of the artificial intelligence industry and the legal order be realized.

This article is from the WeChat official account "Zhichanli" (ID: zhichanli), author: Sui Jingxuan, published with authorization from 36Kr.