Who Owns AI Training Data? The Emerging Legal Battle Between US Fair Use and EU Rights

The rapid growth of generative AI has created a fundamental problem: the technology depends on massive datasets, but the law governing who owns that data, and who may use it, varies sharply across jurisdictions. The core conflict pits existing intellectual property rights against the transformative capabilities of AI, and the world's two largest Western regulatory blocs are answering the question in very different ways. The United States leans on the flexible, case-by-case "fair use" doctrine; the European Union is building a structured, rights-based framework, exemplified by the EU Artificial Intelligence Act, that emphasizes data protection and intellectual property rights.

Introduction: The Evolving Landscape of AI Data Ownership

As AI has gained traction, policymakers and legislators around the world have been grappling with how to regulate it—seeking, as a University of Michigan law course taught by Professor Charles Garrett frames it, to minimize possible harms without sacrificing potential benefits. Because different governments are taking different approaches, the practical significance of any answer depends heavily on where an organization builds or deploys its systems.

What makes the problem especially difficult is that many legal issues remain unsettled. Big questions surrounding intellectual property, copyright, authorship, licensing rights, and remuneration, in Garrett's account, still need to be ironed out. Animated debates within US legal circles revolve specifically around AI ownership. This is not a settled field with clear precedents; it is a landscape in motion, where today's compliance answer may be revised by tomorrow's court ruling or legislative update.

The Core Conflict: Data and Copyright in AI Training

The central tension is straightforward to state and hard to resolve. AI models are trained on existing works—text, images, code, and other data—much of which is protected by copyright or governed by private license terms. Training a model means making copies of those works, at least technically, and the owners of those works increasingly assert that their exclusive rights are being exercised without permission or payment.

An analysis of the governance of training data and copyrights by Marcelo Pasetti and colleagues, published in Discover Artificial Intelligence in 2025, argues that these divergences among national frameworks underscore the inherent challenge of harmonizing copyright regulation for AI training internationally—particularly in balancing three goals that do not always point the same direction: the promotion of innovation, the protection of creators' rights, and the ethical governance of AI development. Any workable framework must somehow serve all three constituencies, and the tension between them necessitates evolving legal interpretations and new regulatory frameworks rather than simple extensions of existing law.

Comparing AI Data Governance: US Fair Use vs. EU Rights-Based Approach

The clearest way to see the regulatory divergence is to place the two models side by side. The comparison below draws on the Springer-published governance analysis by Pasetti and colleagues and the University of Michigan legal course materials.

US and EU approaches to using data for AI training
Dimension	European Union	United States
Core legal principle	Rights-based regulatory framework, exemplified by the EU AI Act, emphasizing data protection and intellectual property rights, with the aim of establishing clear rules for AI systems including their training data.	Fair use doctrine under the US Copyright Act of 1976, addressing limitations on exclusive rights; a flexible, case-by-case defense against infringement claims.
Impact on AI developers	Stricter compliance requirements around data usage, potentially necessitating explicit permissions or adherence to defined data governance standards, within a regulated environment.	A more ambiguous landscape: permissibility depends on a case-by-case fair use analysis, which can create legal uncertainty and potential litigation.
Protection for creators	Generally stronger, requiring consideration of copyright and data protection during training, in an attempt to balance innovation with the rights of data owners and creators.	Determined by how courts interpret fair use in the AI context—still an evolving area, with outcomes often depending on whether the AI's use is considered transformative.

The practical consequence of this comparison is anticipatory: a developer building in the EU should expect defined, rules-based obligations, while a developer building in the US should expect doctrine-driven uncertainty that may only be resolved by litigation. Neither path offers a guarantee, but they fail in different ways—one toward compliance burden, the other toward legal ambiguity.

Beyond Copyright: Ethical and Technical Challenges

Copyright dominates headlines, but it is only one axis of the problem. The Office of the Victorian Information Commissioner, in its guidance on AI and privacy, observes that AI presents social, technological and legal challenges to how we understand and protect information privacy, and argues that AI will require established understandings of personal information to be revisited. A model trained on health records, judicial data, or government service information touches privacy frameworks designed before such reuse was conceivable.

Research on ethical and regulatory challenges of AI in healthcare, published in a 2024 narrative review, identifies a broader set of readiness criteria for data used in advanced systems: data quality, lineage and provenance, extraction and processing errors, traceability, and curation. These criteria matter because training data that cannot be traced cannot be defended, corrected, or responsibly licensed.

Proving Origin: The Challenge of Data Provenance

Data provenance—the ability to document where data came from, under what rights, and through what transformations—sits at the center of making AI governance enforceable. The Pasetti analysis treats provenance as a core technical challenge of generative AI governance precisely because rights and responsibilities cannot be allocated for data whose origin is unknown.

Blockchain is frequently proposed as part of the solution, and a CoinGeek analysis by Gavin Lucas explains its actual scope: blockchains can provide records showing ownership claims, licensing rights, timestamps, and other information, but they cannot decide legal questions such as whether a specific use was fair or infringed upon agreements. The same analysis notes that the underlying copyrighted work and private license terms would not need to be stored on a public ledger. In other words, the technology can prove what happened and when; it cannot substitute for a court determining whether what happened was lawful. Lucas's argument that AI needs to integrate an enterprise blockchain system to ensure data input quality and ownership reflects one advocate's position on the technology's purpose, not its legal effect.

Navigating the Future: Licensing and Responsibility

Licensing is where the technical and legal threads meet. The Pasetti analysis argues that responsibility extends to the adequacy of licenses for datasets and algorithms, addressing the need for a legal environment that supports both open sharing and data privacy. Establishing clear requirements on which datasets should be made publicly available, and under what conditions, is in their view essential to ensuring that only AI systems following ethical and responsible principles are included in the ecosystem.

The analysis also emphasizes that it is essential to distinguish between different types of uses of copyrighted works—training on a work, extracting verbatim passages from it, and producing something that competes with it are not the same activity, and a mature framework will treat them differently. That distinction is one of the key unresolved design questions for any licensing regime.

For its part, the European Commission signaled its direction early: a February 2020 report on the safety and liability implications of AI, the IoT, and robotics stated that a clear and predictable legal framework addressing the technological challenges is required, even while asserting, in accordance with the New Technologies Formation, that in principle existing Union and national liability laws are able to cope with emerging technologies.

Conclusion: Towards Harmonized AI Data Governance

The governing analysis is candid that international harmonization of copyright rules for AI training faces an inherent challenge: jurisdictions are diverging rather than converging, with Brazil, for example, pursuing a hybrid and transitional framework that combines civil law traditions, elements inspired by the EU model, and adaptations to its own socioeconomic and legal context. The tension between intellectual property rights and AI's capabilities is structural, not incidental, and resolving it requires evolving interpretations rather than a single settled answer.

Navigating the Future of AI Data Ownership

Organizations developing or deploying AI systems must proactively assess their data acquisition and usage practices against evolving international legal and ethical standards. Because the US and EU models impose different obligations and different risks, a dataset that is defensible in one jurisdiction may be contested in another, and the fair use analysis that shields an American developer offers no protection in Brussels.

The measurable indicator of progress is concrete: monitor legislative updates, court rulings, and industry best practices in both the EU and US regarding AI training data, particularly concerning copyright and data provenance. This article is not legal advice, and it addresses the data that trains AI models rather than the copyright status of what those models generate—but for anyone building with AI, the divergence between these two regulatory philosophies is the single most consequential fact shaping what data they may lawfully use.