Whether you may use copyrighted works Generative AI Training is a complex and constantly evolving matter. There is no easy answer to this question, and the answer differs based on the jurisdiction and other factors related to the use of works.
In the United States, the central debate revolves around the doctrine of “fair use” and, in particular, whether using copyrighted works for generative AI training is fair use or not.
Those who consider that using copyrighted works for Generative AI training is Fair Use, argue that using copyrighted works to train AI is “transformative” enough and therefore constitutes fair use. They contend that the goal is not to copy the work for the sake of the work itself, but to enable the AI to learn patterns, generate new, non-infringing content, and innovate. In fact , some U.S. courts have held that training AI on copyrighted material can be considered fair use because of this transformative purpose.
The copyright holders on the other side argue that the unauthorized reproduction of their works to train AI models constitutes copyright infringement. They emphasize on the fact that developers often copy entire works of authorship without permission and compensation, which threatens their creative livelihoods and potentially harms the market for their original works.
Currently there are numerous active lawsuits in the U.S. courts, where judges are analyzing whether training models on copyrighted works and material are protected by fair use. Rulings have been mixed, and judges have acknowledged that while some training-related uses might be fair use, others may not be, often depending on the specific facts of the case.
In May 2025, the U.S. Copyright Office released a pre-publication version of the third part of its report on Copyright and Artificial Intelligence, which offers the most comprehensive federal guidance to date on this issue.
Key Points to Remember:
- Possible Infringement: The Copyright Office acknowledges that the training process—which often involves downloading, transferring, and copying massive datasets—implicates exclusive copyright rights, such as the right of reproduction and the right to create derivative works.
- Case-by-Case Analysis: There is no blanket rule. Courts must apply the four statutory fair use factors (listed below) to the specific circumstances of each case, including the nature of the works, how they are used, and the function of the AI system.
- Transformativeness is Variable: While general-purpose models may often be considered transformative, transformativeness is “a matter of degree.” Training AI to generate expressive content that competes directly with the original works is less likely to be deemed fair use than models deployed for non-commercial research or closed, non-substitutive tasks.
- Licensing over Legislation: The Copyright Office favors voluntary, industry-led licensing markets over new statutory or compulsory licensing interventions. The presence of accessible licensing options for specific types of content makes a fair use defense less appropriate for those works.
- Unlawful Access: The Copyright Office takes the view that knowingly using pirated or unlawfully accessed works weighs against a finding of fair use.
The Four Fair Use Factors that should be considered in the context of AI:
- Purpose and Character of the Use: For instance, highly transformative uses (such as research or functional tools) weigh in favor of fair use, while uses that aim to create content competing with the original works disfavor it.
- Nature of the Copyrighted Work: Using highly creative or unpublished works weighs against fair use, whereas using factual or functional material may favor it.
- Amount and Substantiality Used: This factor takes into account the amount of the original work that was used. Because training often requires the use of the entire works, this factor typically weighs against fair use, though the Office suggests courts should also consider whether the AI model effectively prevents the public output of that protected expression.
- The Effect on the Market of the Newly-created Work: This factor is significant; if an AI system disrupts the market for original works or undermines their potential licensing opportunities, it weighs against fair use.
Since this is an evolving matter, if you are considering using copyrighted works to train AI systems, it is essential to monitor developments in your specific jurisdiction and industry. In the meantime, you can consider whether authorization or licensing are a safer pathway for you than relying solely on fair use defenses.
You can also consider implementing guardrails to prevent your AI models from reproducing copyrighted works verbatim, which can help mitigate infringement risks. If you are using third-party AI models or tools, make sure to conduct due diligence regarding their training data sources and copyright compliance practices.
Are you confused about whether your use of AI is copyright infringement? Do you love using AI but need reassurance that you will not be sued? Let us help you. Reach out to Sutter today for guidance. Let us worry about copyrights and you can focus on your business.
