Synthetic Data Inherits, Not Eliminates, Copyright Risk #
- Legal Considerations in the Use of Synthetic Data for AI Development and Finetuning: The Case of LLMs4EU (ACL Anthology) — This academic paper, using the EU’s LLMs4EU project as a case study, argues that synthetic data is not a legal safe harbor but a tool for risk mitigation that requires robust governance. It assesses the legal exposure of synthetic datasets from two angles: the risk that they may still contain infringing content derived from the original training data (a key issue in the
GEMA v. OpenAI ruling on memorized works) and the constraints imposed by the licenses of the models used to generate the synthetic data, which may prohibit using their output to train other models. - AI Training Data Copyright 2026: IP Risks & EU/US Rules (Google Vertex AI Search Result) — Analysis of emerging compliance regimes notes that the EU AI Office’s template for training data summaries explicitly requires disclosure of synthetic data usage. This requirement is based on the legal principle that if a synthetic dataset was generated by a model trained on copyrighted works, the synthetic data inherits the copyright status of the source material, a concept sometimes referred to as “copyright taint.” This formalizes the idea that developers cannot simply launder copyrighted data by passing it through a generative model.
- Anthropic to Pay $1.5 Billion to Authors for Pirated Books (Pasquale Pillitteri) — The final court approval of the $1.5B Anthropic settlement in July 2026 creates a critical legal distinction between the method of acquisition and the act of training. An earlier ruling in the case by Judge William Alsup had suggested that training an AI on copyrighted texts could qualify as fair use. The record-breaking settlement, however, penalizes Anthropic for how it obtained the training data—by downloading books from pirate archives like LibGen. This establishes a precedent that even if training is deemed fair use, the use of illegally sourced datasets is not.
- Generative AI Copyright: Who Owns AI-Generated Content? (Astraea Counsel) — The U.S. Supreme Court’s denial of certiorari in Thaler v. Perlmutter on March 2, 2026, has cemented the “human authorship” requirement as settled law for AI-generated outputs. This decision leaves in place lower court rulings that content generated autonomously by an AI system, without substantial human creative input beyond a simple prompt, is not eligible for copyright protection. This clarifies that while platform Terms of Service may assign “ownership” of outputs to a user, such contractual terms cannot create copyright protection where federal law denies it.
EU Regulatory Framework Moves from Theory to Enforcement #
- EU Plan To Simplify GDPR: Impact on AI and Consent (1on1 SEO) — A proposed “Digital Omnibus” initiative aims to streamline GDPR compliance and is set to significantly alter the rules for AI training data. The proposal would expand the legal basis for processing pseudonymized data under a “legitimate interest” framework. This marks a substantial shift away from restrictive, consent-based interpretations, potentially giving technology companies more flexibility to use European user data for AI development.
- AI Training Data Copyright 2026: IP Risks & EU/US Rules (Google Vertex AI Search Result) — As of 2026, Article 53 of the EU AI Act is now binding law, imposing enforceable copyright documentation obligations on all providers of general-purpose AI models. This is no longer a voluntary framework but a structural compliance condition with extraterritorial scope, affecting US and Asian developers who may not have fully assessed their obligations under the regulation.
The “Right to Be Forgotten” Confronts AI Training Sets #
- The Right to Be Forgotten: Why AI Makes Erasure Technically Impossible — And What We Do About It (DEV Community) — This analysis contrasts the EU’s GDPR-based “right to erasure” with the lack of any comprehensive federal equivalent in the United States for removing data from AI training sets. The most significant US development is the California Delete Act, which creates a centralized data broker registry and a single-portal system for consumers to submit deletion requests. This system becomes operational in 2026 and represents the most ambitious attempt in the U.S. to create a scalable mechanism for data erasure, though its application to AI models remains legally untested.
- Emerging pattern: A clear legal distinction is solidifying between the legality of the input (the provenance and use of training data) and the copyrightability of the output (the generated content). The Anthropic settlement focuses on the former, while the Thaler case settles the latter in the US.
- Emerging theme: The concept of “inherited copyright” or “copyright taint” for synthetic data is becoming a key compliance consideration, moving from academic theory to a principle embedded in regulatory disclosure requirements like those from the EU AI Office.
- Source to watch: The proceedings of computational linguistics and AI conferences, such as the ACL Anthology, are becoming important sources for detailed legal and technical analysis of AI training issues, bridging the gap between pure legal theory and technical implementation.