In May 2025, the U.S. Copyright Office released a report that addresses the increasing tension between tech companies and copyright owners in AI training of copyrighted work. Here are 3 key takeaways:
- AI Learning is not human learning
Many AI companies have made the analogy that AI learning is just like human learning, whereby the machine learns patterns rather than replicate content directly. The analogy is used to justify AI training of copyrighted work under the fair use doctrine. However, the U.S. Copyright Office challenges this analogy. It argues that the learning behavior of human differs significantly from generative AI models. AI's learning involves wholesale ingestion and copying at massive scales, which has traditionally not been legally permissible. Equating AI and human learning oversimplifies the legal and technical distinctions.
Voluntary licensing has become an increasingly active, yet complex pathway for enabling AI training while respecting copyright law (e.g. the HarperCollins AI Licensing Deal). The report highlights barriers to scalability, such as logistical hurdles (e.g., identifying rightsholders), financial burdens for smaller developers, and imbalance in bargaining power. Entities like Collective Management Organizations (CMOs) help streamline licensing by aggregating rights from multiple copyright holders and making bulk licensing more feasible. While promising, voluntary licensing alone may not be sufficient for AI training. Additional flexible licensing models are suggested.
The fair use doctrine is often disputed when it comes to AI training of copyrighted work (text-and-data mining exception in the EU and fair dealing in the UK). AI companies argue that use of copyrighted work is transformative, intermediate, and does not harm the market of the original work. AI models extract and learn general patterns rather than reproducing identical content, thus constituting fair use. However, content creators, publishers, and rights organizations counter this claim by showing examples of AI-generated content that mostly resembles original work. The U.S. Copyright Office does not take a definitive stance and emphasizes the need for case-by-case analysis, considering into account market harm, the nature of use, and availability of licensing options.
Disclaimer: This report is the pre-publication version of the final report, but the final version will not have substantive changes in analysis and conclusion.
