The Legal Test for Fair Use: Data Provenance, Not Training, Is the Crucial Factor #
- AI Training Data Lawsuits: 2026 Case Tracker and Status (Troveo) — A July 2026 analysis of ongoing AI copyright litigation concludes that the central legal risk is turning on data provenance, not the act of training itself. Courts are increasingly separating the acquisition of data from its use in training a model. For example, in the landmark Bartz v. Anthropic case, which resulted in a $1.5 billion settlement, the court found the training process itself could be fair use, but the company was still liable because the training data was sourced from pirate sites like Library Genesis. This emerging legal distinction suggests that while training may be deemed “transformative,” using illegally acquired datasets creates direct liability, a pattern also seen in the remaining claims in the Kadrey v. Meta lawsuit. This framework is further reinforced by a new major lawsuit filed in May 2026 by publishers including Hachette and Elsevier against Meta, which focuses on the sourcing of training data for its Llama models from pirated repositories.
US Copyright Office Solidifies “Human Authorship” Requirement for AI Output #
- AI-Generated Content Copyright: What You Need to Know (FADEL) — In a definitive move, the U.S. Supreme Court on March 2, 2026, declined to hear the case of Thaler v. Perlmutter. This decision upholds the U.S. Copyright Office’s position that works generated by AI without significant human authorship cannot be copyrighted. The ruling solidifies the legal principle that copyright protection extends only to human-created works, leaving purely AI-generated content in the public domain. This has significant implications for companies using generative AI, as they cannot claim ownership over the direct output of models and must rely on the creative contributions of human users to establish copyright.
EU AI Act Moves into Force, Mandating Training Data Transparency #
- AI Act | Shaping Europe’s digital future (European Union) — The EU AI Act officially entered into force on August 1, 2024, with most provisions becoming applicable as of August 2, 2026. A key requirement for providers of General-Purpose AI (GPAI) models, which took effect in August 2025, is the obligation to document and publish a summary of the content used for training. The European Commission has released an official template for this summary, which requires providers to detail the data sources used, including large datasets and top domain names, to enable rights holders to exercise their rights, such as the right to opt-out under the EU’s Copyright Directive. While some compliance deadlines for high-risk systems have been delayed to 2027 and 2028, the core transparency and data governance rules for GPAI models are now active.
- AI Training and Copyright in Europe: A Potential Shift Beyond Territoriality (Potomac Law) — A non-binding but influential resolution adopted by the European Parliament on March 10, 2026, signals a major potential shift in applying copyright law to AI models. The resolution suggests that EU copyright rules should apply to any AI system offered to users in the EU market, regardless of where the model was trained. This challenges the traditional principle of territoriality in copyright law. The report underlying the resolution even proposes a retroactive licensing fee of “5 to 7% of [the] global turnover” to compensate creators. If this approach is adopted in future binding legislation, it would mean that AI models trained in the U.S. or elsewhere on data considered fair use in that jurisdiction could still be liable for infringement under EU law if made available in Europe.
- Emerging theme: A clear distinction is emerging in legal arguments and court rulings between the act of training an AI model and the act of acquiring the training data. Courts seem increasingly willing to consider the former as a potentially transformative fair use, while holding companies strictly liable for the latter if the data was sourced from pirate websites. This “provenance over process” distinction is becoming a central pillar of AI copyright litigation.
- Source to watch: The official European Union website for the AI Act (digital-strategy.ec.europa.eu) is now a primary source for concrete compliance materials, including official templates for documenting training data. This moves the discussion from theoretical legal analysis to practical, mandated action.
- Gap: While there is extensive discussion about opt-out mechanisms, there is a lack of substantive, recent material detailing the technical implementation and adoption rates of these systems. Most analysis focuses on the legal and philosophical arguments against them, rather than the current state of their use in the wild.