AI companies are discovering that the legal rules governing training data are far more complex than they anticipated, with courts in the United States, Germany, and China now establishing concrete boundaries around what data can be used to build AI models. Between 2025 and 2026, landmark cases have emerged that challenge the assumption that scraping data from the internet and using it to train large language models (LLMs) falls safely under fair use protections. The result is a patchwork of compliance requirements that companies must now navigate to avoid costly litigation and regulatory penalties.

What Legal Risks Are AI Developers Actually Facing?

The legal landscape for AI training data has shifted dramatically. A Munich court ruled that OpenAI infringed copyright by using song lyrics for model training without authorization, while the Shanghai Intellectual Property Court adjudicated the first copyright infringement case involving an AI large model in China. These cases signal that courts are no longer treating AI training as a special category exempt from traditional intellectual property law.

The core issue is straightforward: training large AI models requires copying massive volumes of text, images, audio, and video onto servers for preprocessing and model training. Under copyright law in most jurisdictions, this act of reproduction itself may constitute infringement if the developer hasn't obtained authorization from the copyright owner. The problem intensifies because fair use doctrine, which has traditionally allowed limited copying for purposes like research or commentary, may not apply to large-scale commercial AI training. China's Copyright Law, for example, uses a closed-list approach that specifies exactly thirteen circumstances where works can be used without permission. Legal scholars have concluded that large model training "cannot be directly categorized as fair use, nor can it be expansively interpreted as a form of statutory license".

Beyond copyright, companies must also contend with personal data protection laws. A substantial portion of internet-sourced training data contains personal information such as names, contact details, biometric data, and behavioral records. Once this information is incorporated into a training dataset, its use is regulated by personal information protection laws like China's Personal Information Protection Law (PIPL). The challenge is that obtaining individual consent from every person whose data appears in massive datasets is practically infeasible, yet the legal standards for processing publicly disclosed personal information remain unclear.

How Are Companies Building Compliant Training Data Strategies?

Rather than relying on fair use arguments, legal experts recommend that AI developers adopt a multi-layered compliance framework. The most straightforward approach is licensing and authorization, where developers obtain lawful rights to use training data by entering into licensing agreements with copyright owners or copyright collective management organizations. However, this approach faces practical challenges, including high licensing costs, fragmented ownership of rights, and low negotiation efficiency.

A second strategy involves prioritizing open-source and public-domain data. Developers can 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. This approach requires upfront investment in data curation but provides greater legal certainty than relying on fair use arguments.

The third pathway involves technical compliance measures. During data preprocessing, developers can screen training data for copyright-related risks and avoid using content expressly marked as prohibited for AI training. Technical measures can 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.

For personal data protection, companies should conduct compliance reviews of data sources and refrain from using data obtained from unlawful sources. During data preprocessing, personal information should be anonymized or de-identified so that individuals cannot be identified from the dataset. Some legal scholars have suggested that the "purpose of disclosure" for publicly disclosed personal information may be interpreted relatively broadly 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.

Steps to Protect Your AI Service Agreements from IP Liability

• Clarify Data Ownership: Service agreements should explicitly state who owns inputs (data entered into the tool) and outputs (content generated by the AI). OpenAI's approach, for example, assigns all ownership rights over inputs and outputs to the client, while reserving the right to use content to improve services. Your agreement should reflect your specific business model and risk tolerance.
• Address Third-Party IP Rights: Incorporate licensing terms, usage rights, indemnities, and warranties related to third-party training data. The agreement should make clear that clients are responsible for ensuring they have the right to post any content they submit, and that they will comply with any third-party licenses relating to that content.
• Include Robust Indemnification Clauses: Require clients to indemnify and hold your company harmless from claims arising from violations of third-party rights, including copyright, property, and privacy rights. This shifts the financial burden of defending against infringement claims to the party best positioned to control the risk.
• Implement Liability Limitations: Use limitation of liability provisions to specify what types of damages your company is responsible for and set caps on total liability. This reduces your financial exposure in the event of an infringement claim.
• Account for Legal Uncertainty: Acknowledge in your agreement that the law around AI and intellectual property is still developing. This sets realistic expectations with clients and may provide some protection if legal standards shift after the agreement is signed.

The challenge for service providers is that the law has not yet caught up to the rapid development of AI technology. Businesses are deploying AI tools even as the legal landscape around intellectual property continues to shift. This uncertainty makes it critical for companies to have a firm grasp of the IP-related terms in their agreements and to update those terms regularly as case law develops.

What Do AI Disputes Actually Look Like in Court?

A comprehensive analysis of more than 425 publicly filed AI-related disputes reveals that the litigation landscape has expanded far beyond copyright claims. While intellectual property matters represent the largest issue cluster at approximately 47 percent of disputes, the docket now includes civil rights and privacy claims (about 15 percent), financial services disputes (about 9 percent), criminal justice and law enforcement cases (about 8 percent), and employment litigation (about 7 percent).

The rate of new AI-related filings has accelerated dramatically. The number of filings climbed from 34 in 2022 to 66 in 2023, 62 in 2024, and roughly 85 in 2025, representing a single-year increase of approximately 35 percent. Generative AI accounts for the largest share of disputes at roughly 59 percent, but the docket now includes AI-enabled decision systems (about 15 percent), computer-vision applications (about 11 percent), predictive and machine-learning models (about 9 percent), and autonomous systems and robotics (about 5 percent).

A structural feature of AI litigation that complicates expert analysis is that the question of what an AI system did and the question of what that conduct is worth are often not readily separable. Both turn on the same underlying technical facts. This interdependence blurs the conventional line between technical and damages issues, straining the sequential division of expert labor on which litigation ordinarily relies. In roughly one in six of the studied matters, the pattern involves large-scale acquisition through scraping and the circumvention of platform protections, including allegations that developers bypassed technological protection measures in violation of the anti-circumvention provisions of the Digital Millennium Copyright Act (DMCA).

The concentration of disputes among a small number of model developers means that forensic methodology developed against those architectures, such as log reconstruction, prompt replay, and version comparison, can be reused across matters rather than rebuilt case by case. Federal district courts exercising original jurisdiction handle the clear majority of AI-related matters, well outpacing the combined volume of state trial courts and foreign courts, though the meaningful minority of matters proceeding in state and foreign venues means experts cannot assume a single governing standard for admissibility and method will apply across every case.

As AI systems become more deeply embedded in business operations and consumer products, the disputes surrounding them will likely continue to expand. The most consequential AI disputes of the coming decade may arise less from what these systems create than from how the organizations deploying them choose to govern their use, including questions of oversight, controls, and risk management.

" }