[Direct answer]Is using others' works to train artificial intelligence models an infringement? This requires distinguishing between the training process and the output results and judging them separately. Article 24 of the Copyright Law of the People's Republic of China (effective June 1, 2021) adopts a closed-list approach to fair use, and model training is not explicitly included. Unauthorized large-scale collection of language data poses a risk of infringement. If the model's output constitutes substantial similarity to the original work and encroaches on the original work's market, the rights holder may claim to stop the infringement and seek compensation, with the statutory compensation range being between 500 yuan and 5 million yuan (Article 54 of the Copyright Law). Enterprises should control the risk by obtaining authorizations and retaining authorization documents.
1. Problem background and typical scenarios
The training of large-scale models and image generation tools relies on massive corpus data, with online novels, news, photographic images, illustrations, and code being common objects of data collection. When creators discover that their works appear in publicly available training data lists, or that the models can generate content highly similar to their works, their first reaction is often that there's no way to protect their rights: the works have already spread across the internet, and it's difficult to identify the actual data collectors and users. Platforms often defend themselves by claiming technological neutrality, so the starting point of protecting rights often gets stuck on the verification of evidence and the identification of the responsible parties.
Companies also have concerns. When the algorithm team collects data from the public network, they usually only pay attention to the robots protocol and access frequency, ignoring whether the copyright authorization chain is complete. When fine-tuning the model with customer data or employees' works, it may also trigger confidentiality and ownership clauses with customers and partners. In recent years, intellectual property courts in Beijing, Shanghai and other places have accepted a number of disputes involving training data, and the adjudication guidelines are still being formed. It is far more cost-effective to comply with regulations in advance than to defend against lawsuits afterwards.
There is another high-incidence scenario involving internal use within enterprises: the marketing department uses generation tools to produce copy that is highly similar to competitors' advertising slogans, or the design team directly applies style data collected online to client project deliveries. In such cases, the responsibility for infringement is typically borne by the enterprise. The distinction between employees' personal actions and their official duties, as well as the recourse relationship between the enterprise and data service providers, will all be scrutinized when disputes arise. The completeness of the contractual provisions often determines who ultimately bears the costs.
II. Key points of dispute
One of the focuses is the qualification of the training behavior itself. Supporters of fair use emphasize the value of transformative use and technological innovation, while opponents point out that commercial training constitutes a substitute for the original work market. Article 24 of the current Copyright Law does not list machine learning scenarios, so the final conclusion can only be left to case-by-case adjudication, and the standards of different courts have not yet been fully unified.
The second focus is on the determination of infringement of the output results. If the content generated by the model is substantially similar to the original work and the model has the ability to access the original work, it may constitute infringement of copying or adaptation; if it is only similar in style and remains at the level of conceptual borrowing, it generally does not fall within the scope of exclusive rights. The third focus is on the contract and subject level. Whether the authorization of data suppliers is authentic and complete, and how the contagious clauses of open-source model licenses constrain downstream use, all directly affect the assumption of responsibility and the path of seeking compensation.
The technical details at the algorithmic and factual levels also affect the arbitration: whether the training set includes complete works or excerpts, how many recognizable expressions remain in the original works after deduplication and cleaning, and the degree of overlap between the model's output and the original work in terms of probability distribution will all be analyzed item by item by the experts on both sides. If the rights holders only submit general similarity screenshots without conducting quantitative comparisons, their claims are often difficult to be fully accepted by the court, a point that has been repeatedly highlighted in existing precedents.
III. Professional analysis by Hengluo Lawyers
The determination of liability and reduction of compensation in intellectual property cases highly depend on the organization of evidence and the application of law. Taking a trademark infringement and unfair competition dispute handled by our firm as an example: the plaintiff, Company A, is a well-known paper industry enterprise. The "EFGH" trademark is a well-known trademark for sanitary napkins in Category 5. The defendant, Company B, obtained authorization from a third party in September 2023 and used the "EFGK" trademark and design patent to produce and sell "EFGK" sanitary napkins, which were then sold through the e-commerce channel of Company C. In April 2024, Company A filed a lawsuit, alleging that Company B imitated the well-known trademark and confused the packaging and decoration, seeking compensation of 980,000 yuan. Company B appointed Attorney Liu Yuanyuan from Beijing Hengliu Law Firm to defend the case. The attorney raised defenses such as the logo having legal authorization and not constituting similarity, the packaging and decoration not falling into unfair competition, and the high compensation lacking basis.
After the court's trial, it partially adopted the lawyer's opinion: on the determination of infringement, "EFGH" had already constituted a well-known trademark when "EFGK" was applied for, and the use of "EFGK" constituted imitation, so the trademark infringement was established; but Company A failed to prove that its packaging and decoration had a certain influence, and its claim of unfair competition was rejected. On the liability, the court ordered Company B to stop the infringement and publish a notice to eliminate the impact. Given the short infringement period and small business scale, the court ruled to compensate for economic losses of 50,000 yuan and reasonable expenses of 10,000 yuan, totaling 60,000 yuan. This case was represented by Attorney Liu Yuanyuan (practice certificate number: 11101202011201094, subject to the certificate issued by the judicial administrative organ).
The direct implication of this case for AI training disputes is that the scope of rights and the quality of evidence determine the amount of compensation. The plaintiff lost a claim due to insufficient evidence of packaging and decoration, while the defendant was unable to completely avoid liability due to the failure to establish its well-known trademark. In cases involving training data, it is also necessary to do a thorough job on the authorization chain, the possibility of exposure, and the comparison of substantial similarity. Both parties should prepare sufficient hard evidence, and the final outcome can vary by several times.
Regarding the training data dispute itself, Henglu Law Firm advises focusing on three key points: First, contact and similarity are the two pillars for establishing infringement. It is essential to compare the training set list with the output results. Second, the calculation of compensation should start from the logic of commercial substitution, explaining which licensed markets have been occupied by the infringing content. Third, settlement is often a more cost-effective solution. Exchanging licensing fees for long-term compliant use is far more beneficial to the continuity of both parties' businesses than protracted litigation.
IV. Practical suggestions and legal basis
The company recommends implementing four compliance measures: First, sort out the sources of training data, prioritize purchasing commercial corpora with clear authorizations, and keep records of the scope and time of data collection when collecting data from the public internet; Second, sign authorization agreements with data suppliers to clarify ownership guarantees and breach of contract responsibilities; Third, when using customer data in fine-tuning processes, obtain written consent and stipulate model ownership in contracts; Fourth, conduct similarity checks on output content before launching the model, establish complaint response and removal mechanisms, and keep records of processing results.
The rights holder's path to protecting their rights is as follows: First, they fix the core evidence such as the timestamp, blockchain evidence, or notarization of the training set list and the comparison results of the model output; then, they send a notice of infringement to the platform or the model operator, negotiate the licensing fee or request the removal of the content; if unsuccessful, they file a lawsuit with the intellectual property court or intermediate court with jurisdiction, and may claim to stop the infringement, eliminate the impact, and seek compensation. The amount of compensation shall be calculated based on the actual losses of the rights holder or the infringer's profits. If it is difficult to determine the amount of compensation, the statutory compensation provided for in Article 54 of the Copyright Law shall apply. The right holder must assert their rights within the three-year statute of limitations stipulated in Article 188 of the Civil Code of the People's Republic of China (effective from January 1, 2021).
In terms of the supporting system construction, it is recommended that enterprises appoint data compliance officers and keep records of the four stages of data collection, cleaning, storage, and training respectively: - In the collection stage, record the sources and robots restrictions. - In the cleaning stage, record the deduplication rules. - In the storage stage, register the corpus fingerprints. - In the training stage, lock the corresponding relationships between versions and datasets. This set of records is typically used for compliance self-inspections. In the event of disputes, it serves as the most powerful evidence of exoneration, with the cost being just a process document and a small amount of working hours.
5. Henglu Law Firm's Advice
The compliance window for training data is narrowing. The earlier companies establish data registries and authorization files, the more smoothly they can navigate subsequent financing reviews, due diligence, and litigation defense. For creators, they shouldn't give up on protecting their rights just because the value of a single work is limited. The combined cost of batch registration and batch storage is manageable, and the bargaining power will also increase accordingly. For companies that frequently produce content, they can also uniformly mark copyright information and storage numbers when releasing content externally, so that each work has a clear ownership fingerprint.
Both enterprises and creators can conduct a written assessment when encountering disputes, clearly listing the authorization gaps, evidence shortcomings, and room for settlement, before deciding whether to negotiate or sue. For more services, please refer to the following link:Intellectual property lawyerColumn.
VI. Frequently Asked Questions
Q1: Is it considered fair use to train a model using publicly available corpus collected from the internet?
The current law does not provide a direct affirmative answer. Article 24 of the Copyright Law of the People's Republic of China (effective June 1, 2021) adopts a closed-list approach, and situations such as personal learning, classroom teaching, and scientific research do not cover commercial model training. If captured language materials are used for commercial large-scale models, there is a real risk of being deemed infringement. Enterprises should obtain legal sources through purchasing authorized language materials and signing data authorization agreements, and keep the authorization documents for future reference. They cannot evade their obligation of due diligence solely on the grounds of technological neutrality. In judicial practice, courts typically evaluate cases based on the purpose, market scale, and impact on the original work's market. Once the case crosses the threshold of commerciality, the scope for defense has significantly narrowed.
Q2: If a rights holder alleges that AI training infringes their rights, what evidence should they provide?
The core evidence can be divided into three layers: 1. **Ownership layer**: Submit the work registration certificate, the original draft of the work, or the record of the first publication. 2. **Behavioral layer**: Use time stamps or notarization to fix the list of training data, capture logs, and the fact that the authorization chain is missing. 3. **Result layer**: Collect samples to compare the content generated by the model with the original work, and if necessary, apply for a judicial appraisal to conduct a substantive similarity analysis. The more complete the evidence is in demonstrating contact and similarity, the easier it is to gain support for the claim, and the higher the potential for compensation. For models involving overseas entities, it is also necessary to cross-check the training data disclosed by the other party with the recorded information to identify the responsible parties.
Q3: Approximately how much would it cost to sue an AI company to protect one's rights?
The acceptance fees for civil intellectual property cases are calculated according to the amount of the subject matter in stages, as stipulated in the "Measures for the Payment of Litigation Costs". Those claiming statutory compensation shall pay according to the amount of the claim, and the losing party shall bear the fees or share them according to the proportion of the victory and defeat. Lawyer's fees are charged on a per-case, hourly, or percentage basis of the subject matter. Evidence preservation and appraisal fees usually range from a few thousand yuan to tens of thousands of yuan. It is recommended to first assess the strength of the output comparison results before determining the amount of the claim, so as to avoid paying excessive acceptance fees for overly high claims. It is also possible to resolve disputes through authorization negotiations before litigation. The losing party is usually required to bear the reasonable expenses incurred by the right holder to stop infringement, including the notarization fees and lawyer's fees that are supported.
Q4: What are the procedural milestones that an intellectual property case goes through from the filing to the judgment?
The general process is as follows: Submit the complaint and ownership evidence to complete the case filing. The defendant submits a defense within 15 days of the defense period; both parties present evidence and cross-examine each other, and arrange for judicial appraisal involving technical comparison. Subsequently, the court holds a hearing, and the first-instance judgment is usually made within six months after the case is filed, and the case is concluded within three months if the summary procedure is applied. If the defendant is not satisfied with the first-instance judgment, he/she can appeal within 15 days, and the second-instance judgment is generally concluded within three months. The Beijing Intellectual Property Court and intellectual property courts in various regions handle such cases with relatively professional procedures and faster trial pace. For cases involving protective measures or technical investigations, the court may intervene in advance upon request, which will correspondingly prolong the overall timeframe.
Q5: If I find that my work has been used to train a model, what should I do first?
First, secure the evidence before contacting the other party. Step 1: Take screenshots of the consistency between the training data list, the model generation results, and your own works, and preserve them with a trusted timestamp or notarization; Step 2: Inquire about the main information and contact details of the model operator; Step 3: Send a written infringement notice and retain the delivery certificate, demanding that they stop using the model and provide an explanation of the authorization source; Step 4: Based on the response, choose to negotiate authorization or initiate litigation, paying attention to the calculation and interruption of the three-year statute of limitations throughout the process. For bulk works, you can create a rights list according to the same template and refer to it directly when negotiating with the platform or initiating litigation.
The above Q&A is mainly based on judicial practice in the Beijing area. There may be differences in the adjudication guidelines in different regions. It is recommended to consult a professional lawyer based on the specific circumstances of the case.
If you are currently facing similar problems, it is recommended that you consult a professional lawyer as soon as possible with the relevant evidence materials, to avoid missing the statutory time limit. →Intellectual property protection legal services·Attorney Liu Yuanyuan·Beijing Henglu Law Firm
Disclaimer: This article is compiled based on current effective laws and regulations and public adjudication rules, and is for general reference only. It does not constitute a formal legal opinion on specific cases. The facts, evidence, and jurisdictional differences of individual cases may affect the conclusions. In the event of a dispute, please consult a practicing lawyer in a timely manner.