Can copyright survive AI? The legal storm that’s changing everything

Today, copyright law is confronting one of its biggest challenges yet: the rise of generative artificial intelligence (AI). Generative AI refers to systems that can create content — such as text, images, music, or code — largely on their own, often producing results indistinguishable from human-made work. This rapid advancement has outpaced the evolution of existing law. Legal institutions worldwide are struggling to make sense of how doctrines crafted for human authors apply to machine-generated output. In the absence of clear rules, AI is testing the boundaries of creative industries at an unprecedented speed, upending traditional notions of authorship, ownership, and originality

Two key questions dominate global discussions:

1. Can companies freely use copyrighted material to train AI models, or does this constitute an infringement of copyright?
2. Is output generated by AI, based on a human prompt, eligible for copyright protection itself?

The image in the header attached to this article was generated by AI. As many companies worldwide begin incorporating AI into their daily operations, several questions become increasingly relevant:

• Is the company providing the AI-generated image liable for copyright infringement (should it have compensated the original content owners for training the AI), and
• Is this image protected under copyright law?
• Can an individual or a company sue a third party for using content that they generated through AI?

As is often the case, the answers to these questions are nuanced and significantly dependent on jurisdiction.

In this first article, we will focus exclusively on examining the question of copyright infringement associated with AI model training across various jurisdictions. A subsequent article will discuss the second question — whether AI-generated outputs based on human prompts are eligible for copyright protection as of March 2025.

Let’s review how the legislative bodies and courts address this question in several key geographies.

1. United States

A recent ruling on February 11, 2025, by the U.S. District Court of Delaware in Thomson Reuters v. Ross Intelligence marked a significant development in copyright law concerning AI.

• The court ruled that using copyrighted materials to train competing AIproducts does not constitute fair use, thus holding Ross Intelligence liable for copyright infringement.
• A notable aspect of this ruling was the rejection of the previously accepted defence (as seen in Google v. Oracle, Sony v. Connectix, and Sega v. Accolade), where a fair use exception was recognised because AI training was considered an “intermediate step.”
• The court distinguished these precedents, emphasising that those cases involved copying computer code, which serves a functional purpose, unlike books, films, and other literary works.

The court highlighted that “undoubtedly the single most important element of fair use” it considered was the likely impact of the copying on the market for the original work. Significantly, this included not only the current markets but also derivative ones “that creators of original works would generally develop or license others to develop” — in other words, the market for data licensing to train AI. This ruling expands the criteria for assessing market substitution effects (Factor 4 of fair use) to determine if the copied work’s value or potential market was impacted, covering not only immediate markets but also licensing markets specifically for AI training.

2. China

As of March 2025, China’s regulatory and judicial approaches to AI training using copyrighted materials reflect a nuanced effort to balance technological innovation with intellectual property (IP) protection.

• Although definitive legislation has yet to address explicitly the legality of using copyrighted works in AI training, both regulatory developments and judicial precedents provide significant guidance.
• Draft proposals under public consultation suggest a potential “reasonable use” exception for AI training, permitting copyrighted materials if their use diverges significantly from the original work’s function and does not cause unreasonable harm to rights holders.
• Concurrently, mandatory AI labelling requirements (effective from March 2025) aim to enhance transparency for AI-generated outputs, indirectly encouraging developers to audit their training data sources.

Notable recent cases include:

• February 2024, the Guangzhou Internet Court rendered China’s first ruling concerning AI-generated content in the Ultraman case. The facts differed slightly as the defendant was a platform offering generative AI services based on user-provided inputs. The platform itself did not directly train the AI models but provided the means to do so. The court concluded that the platform was liable for copyright infringement, noting substantial similarities between AI-generated outputs and pre-existing copyrighted works. However, the court refrained from ruling on whether training data usage itself constitutes infringement.
• Another pending lawsuit involves Xiaohongshu’s Trik AI application, challenging the use of illustrators’ works for model training under China’s Copyright Law. This case could set significant precedents for future disputes.

China’s approach currently priorities accountability for AI-generated outputs and platform governance rather than directly scrutinising the use of training data. Although clear judicial rules are still evolving, pending legislation and judicial trends indicate increasing compliance burdens for developers. Companies are encouraged to implement safeguards, such as copyright recognition tools, and pursue licensing agreements for high-risk datasets as they navigate this evolving regulatory environment.

3. United Kingdom

The UK government is actively addressing this question through a consultation process launched back in December 2024. Currently, the UK’s copyright framework is considered restrictive, placing it at a competitive disadvantage. The government is contemplating several options:

• Maintaining the status quo (leaving copyright laws unchanged);
• Strengthening copyright protection by requiring explicit licensing for all AI training data, significantly impacting the UK’s AI sector;
• Introducing a broad data mining exception to allow AI developers unrestricted use of copyrighted materials, although this faces opposition from creative industries.

The most favoured proposal mirrors the European Union’s (EU) approach, introducing a data mining exception allowing rights holders to opt out. This would grant AI developers access to high-quality training data while enabling rights holders to control and receive remuneration for their work, complemented by enhanced transparency measures.

4. European Union

The EU has established a clear legislative framework through the AI Act and the 2019 Copyright Directive, emphasising strict compliance with copyright law.

• By Q3 2027, AI developers must document training data sources and transparently disclose copyrighted materials used.
• Significantly, the Copyright Directive introduces a text and data mining (TDM) exception for commercial purposes, allowing AI developers to train models unless rights holders explicitly prohibit it via machine-readable means.

The draft Code of Practice further elaborates on compliance expectations, stressing “reasonable measures”, including making “best efforts” to avoid infringing outputs and promptly addressing rights holder complaints. However, its ambiguous language creates uncertainty for developers.

In the notable Kneschke v. LAION case in Germany (the EU’s first major test of AI training legality), the Hamburg Regional Court ruled that LAION’s downloading of copyrighted images to create datasets fell within the scientific research exception under Section 60d of the German Copyright Act. This provision permits non-commercial research entities to reproduce copyrighted works for TDM purposes. The court emphasized LAION’s non-profit status and the free public availability of its datasets. However, it also underscored the necessity for rights holders to use effective machine-readable reservations to block TDM under commercial exceptions.

As the exploration of AI and copyright law has shown, the uncertainty in this area is far from theoretical — it carries real-world consequences for individuals, businesses, and entire creative industries. Creative professionals are already feeling the effects. While generative AI opens exciting new opportunities, it also raises the spectre of artistic displacement and unfair competition. Many artists, writers, and other creators worry about losing control over their work or seeing their livelihoods eroded by AI-driven substitutes. These concerns are not merely speculative. In some instances, people have started to lose work, finding themselves unexpectedly competing with AI-generated imitations of their own style or even their own likeness.

Facing these high stakes, policymakers and stakeholders are urged to proceed with caution — but also with purpose. Crafting a balanced approach to AI and copyright will be crucial to navigate the path ahead. On one hand, we must uphold the core principle of intellectual property law: rewarding human creativity and protecting original expression so that artists and authors can continue to make a living from their craft. On the other hand, we cannot allow rigid or outdated laws to stifle technological progress and the benefits it promises. The central challenge, therefore, is finding a compromise that safeguards creators’ rights without smothering innovation.