Who Owns AI-Generated Content? Copyright Law in the Age of Generative AI

May 29
Copyright is the most legally contested area of generative AI in 2026, and most coverage collapses three different questions into one. This post separates them — who owns the training data, who owns the output, and what the EU AI Act now adds on top through Article 53's training-data opt-out and Article 50's labelling obligation. It also maps the meaningful differences between the EU, UK, and US, including the UK's distinctive section 9(3) carve-out for computer-generated works. Useful for in-house counsel, content leads, and IP advisers building defensible generative-AI workflows.
Copyright is the most legally contested area of generative AI in 2026. Whenever a marketing team drafts a blog with ChatGPT, designers generate hero images with Midjourney, or in-house lawyers fine-tune a model on client data, three different IP questions are in play at once. Pulling them apart is the first step to clear thinking.
In the case of distance learning, it may be most appropriate at colleges and universities. Research data consistently indicate that students strongly prefer distance education.

The most consequential battleground is training data. Modern foundation models ingest trillions of words and billions of images, most of them copyrighted. Whether that ingestion is lawful is the unresolved question. The EU's answer rests on the text-and-data-mining regime in Article 4 of the DSM Copyright Directive — mining is permitted unless rights-holders have explicitly opted out. The EU AI Act reinforces this through Article 53, which requires general-purpose AI providers to identify and respect those opt-outs even for training that happened outside the EU, and to publish a public summary of training content using the AI Office's mandatory template. The US position is messier, with ongoing fair-use litigation in cases like NYT v. OpenAI, Authors Guild v. OpenAI, and Getty v. Stability AI. The UK declined to legislate in its March 2026 report, leaving rights-holders dependent on common-law claims.

The output question — and why human creativity still matters

The question most clients actually ask is whether they can own what the AI produces. The answer turns on how much a human shaped the result. In the US, the Copyright Office has been consistent: copyright protects only human authorship. Purely AI-generated images receive no protection (Théâtre D'opéra Spatial, 2022); AI-assisted works receive thin protection only over the human-curated portions (Zarya of the Dawn, 2023). The EU position is similar in principle, applying the CJEU's existing originality test — the work must reflect the author's "free and creative choices." The UK is the exception. Section 9(3) of the Copyright, Designs and Patents Act 1988 attributes authorship of "computer-generated works" to the person who made the arrangements necessary for the work's creation — a provision drafted in 1988 for video-game graphics that has unexpectedly become a global outlier in the AI era.

The practical implication for businesses is that human curation is what creates protectable IP. A marketing post substantially rewritten by a human author carries thin but real copyright in those edits; a pure prompt-and-publish output usually does not. From August 2026, Article 50 of the AI Act will also require providers of generative AI to mark outputs in a machine-readable format and deployers to disclose deepfake content — adding a transparency layer on top of the underlying IP analysis. The honest summary: assume AI outputs may be unprotectable, training inputs may infringe, and the workflow design — not the choice of tool — is what creates value you can defend.

Learn more about the EU AI Act on Lexstream's AI Law and Governance Series.