Artificial Intelligence (AI) is changing the game across industries, but it also brings a major headache for intellectual property (IP) laws, especially copyright. The European Union’s AI Act steps in to regulate AI systems, but it raises serious questions about how copyrighted works are used in AI training. Where do we draw the line between fair use and infringement? What rights do content creators actually have when their work gets absorbed into massive AI datasets?
With high-profile lawsuits like The New York Times v. Open AI and Microsoft happening in the U.S., these issues are more pressing than ever. In this piece, we’ll break down how the AI Act impacts copyright, text and data mining (TDM) exceptions, and the challenges of enforcing these rules in a world where AI models process untraceable amounts of data.
AI TRAINING AND COPYRIGHT – WHERE’S THE LINE?
AI systems rely on huge amounts of data to learn and generate new content, but that data often includes copyrighted material. This raises the big legal question: Does using copyrighted works to train AI count as infringement?
Under Article 53 of the AI Act, there’s now a direct link between training data and AI models, putting AI-generated content under closer legal scrutiny. Right now, it’s incredibly difficult to trace whether an AI platform has used protected content, making enforcement tricky.
At its core, AI training involves analyzing vast datasets to recognize patterns, improve performance, and generate content that mimics human creativity. But the key issue here is that AI models don’t just analyze data—they learn from it, often incorporating elements of copyrighted materials into their output. This means that while an AI system may not explicitly reproduce a copyrighted work, it still benefits from it, leading to serious concerns among copyright holders.
Differing Legal Opinions: Is AI Training Fair Use or Infringement?
There are two dominant schools of thought regarding whether AI training on copyrighted content constitutes infringement:
The Pro-AI Argument (Fair Use Perspective):AI advocates argue that training models don’t actually “reproduce” works; instead, they analyze patterns, extract statistical information, and generate something new. From this perspective, AI training could be compared to a human reading thousands of books and then writing something original based on learned knowledge. Some legal scholars liken this to the fair use doctrine, suggesting that AI-generated outputs are transformative and do not compete with original works.
· The Copyright Holders’ Argument (Unauthorized Reproduction):On the other side, content creators and copyright holders argue that AI is essentially profiting from their work without permission. Since AI models generate outputs based on existing copyrighted materials, they believe this constitutes unauthorized reproduction, even if the output isn’t an exact copy. Many point out that AI-generated works can imitate the styles of specific artists or authors, which suggests a strong dependence on copyrighted material.
Challenges in Enforcing Copyright Protections
The biggest roadblock to enforcing copyright protections against AI training is the lack of transparencyin how AI models are trained. Many AI companies refuse to disclose their training data, citing trade secrets and competitive advantage. This makes it nearly impossible for copyright holders to know whether their works have been used.
Additionally, AI-generated content is often non-traceable—unlike traditional copyright violations where an infringing work is an exact or near-exact copy, AI-generated works may only retain fragments of copyrighted material, making infringement harder to prove.
One potential legal solution is the burden of proof reversal, which some policymakers suggest: AI developers should be required to demonstrate that their models were trained on legally acquired data. However, this approach faces pushback from AI companies that fear disclosing training methodologies could undermine their business models.
So far, the legal system hasn’t given us a clear-cut answer. The EU’s Directive 2019/790tried to clarify things by introducing exceptions for text and data mining (TDM), which we’ll get into next.
TEXT AND DATA MINING (TDM) – COPYRIGHT EXCEPTIONS AND OPT-OUTS
One of the biggest legal loopholes for AI training is the TDM exception, which allows the use of copyrighted material for certain purposes. The EU Copyright Directive carves out two key exceptions:
· Article 3: Allows TDM for scientific research, meaning researchers can freely mine data unless expressly prohibited.
· Article 4: Allows TDM for any other purpose, as long as the copyright holder hasn’t opted out.
However, this has led to an evolving legal landscape where AI developers, content creators, and policymakers are struggling to define the boundaries of lawful data usage. AI models rely on vast amounts of data to improve performance, and often, this data includes copyrighted works. While the AI Act acknowledges the necessity of regulating this practice, it does not provide a foolproof framework for managing the rights of copyright holders in a fair and enforceable way.
The Opt-Out Mechanism – A Legal Grey Zone
While copyright law technically grants rightsholders the ability to opt out of AI training, the real-world application of this right is riddled with challenges. The current system lacks clarity, consistency, and enforcement, leaving copyright holders in a precarious position.
1. The Lack of a Standardized Opt-Out System
There is no single, universally recognized method for rightsholders to declare their refusal. Should they send individual takedown notices? Embed metadata in their works? The law offers no definitive answer. This ambiguity benefits AI developers, who can argue that without a clear, standardized opt-out mechanism, enforcement remains impractical.
2. The Hamburg Court Decision – A Weak Precedent
In a recent ruling, the Hamburg court held that an opt-out could be expressed in “natural language.” While this might seem like a step toward flexibility, it ultimately introduces more uncertainty. What qualifies as a valid declaration? How will AI companies interpret these statements? The ruling lacks specificity, paving the way for inconsistent enforcement across jurisdictions.
3. Central Registries and CMOs as Potential Solutions
To address these issues, some experts have proposed:
· A centralized registry, as suggested in a French report on AI and copyright, where rightsholders could log their works to prevent unauthorized AI use. However, such a system would require international cooperation and extensive administrative oversight.
· Greater involvement of Collective Management Organizations (CMOs), which already help artists manage copyright licensing. By acting as intermediaries, CMOs could negotiate AI training agreements, monitor compliance, and take legal action against violations.
4. Loopholes and AI Non-Compliance
Even if an opt-out is declared, enforcement remains a major hurdle. AI developers process vast amounts of data, and some works might still be scraped—intentionally or inadvertently. Without strict auditing requirements or penalties for non-compliance, the opt-out mechanism remains more theoretical than practical.
5. The Challenge of Cross-Border Enforcement
AI training isn’t confined to a single jurisdiction. A work opted out in the EU could still be used in AI training conducted elsewhere, raising significant concerns about enforcement. Without international alignment on copyright exceptions and AI governance, rightsholders may find themselves fighting an uphill battle.
The current opt-out framework is riddled with loopholes, inconsistencies, and enforcement gaps. Until clearer legal standards, stronger oversight mechanisms, and global cooperation emerge, the battle between AI innovation and copyright protection will remain unresolved.
TRADE SECRETS VS. TRANSPARENCY – THE BALANCING ACT
One of the biggest challenges in AI regulation is striking a balance between protecting trade secrets and ensuring transparency. AI companies invest heavily in developing their models and argue that their training datasets, algorithms, and methodologies are proprietary assets that give them a competitive edge. Requiring full disclosure of these datasets could expose them to intellectual property risks and undermine their business models.
On the other hand, transparency is a growing legal and ethical concern, especially when AI systems use copyrighted content without authorization. Without clear information on what data is being used and how models are trained, copyright holders have no way of knowing whether their work has been exploited, raising serious legal questions. Transparency is also essential for accountability—AI-generated outputs can sometimes reflect biases or inaccuracies, and understanding the underlying data sources is key to addressing these issues.
The EU AI Act attempts to navigate this delicate balance by requiring AI developers to disclose certain details about their training datasets while still allowing for the protection of trade secrets. However, the challenge lies in how these rules will be implemented across different EU countries. Copyright laws and trade secret protections can vary, and if individual member states interpret these rules differently, AI companies could find themselves navigating a complex patchwork of regulations.
This lack of harmonization could create uncertainty for businesses, discouraging AI innovation within the EU and pushing companies to operate in jurisdictions with more predictable legal frameworks. As a result, there is mounting pressure for a unified EU-wide approach to AI and intellectual property law—one that ensures transparency without stifling innovation or compromising proprietary interests.
Where Do We Go from Here?
The AI Act is a step in the right direction, but it’s far from a perfect fix. Copyright holders still lack clear and effective opt-out mechanisms, and the legal debate over whether AI training constitutes copyright infringement remains unresolved. Meanwhile, AI developers are navigating the tricky balance between transparency and innovation.
A few key takeaways from the current landscape:
1. AI training exists in a legal gray zone, and courts will play a crucial role in setting boundaries.
2. While text and data mining (TDM) exceptions allow for some AI training, opting out is still inconsistent and varies across the EU.
3. Stronger enforcement—such as a centralized registry or greater involvement from collective management organizations (CMOs)—could help strike a fairer balance between protecting creators and fostering AI development.
As AI technology rapidly advances, the laws governing it need to keep up. But will legislation evolve fast enough, or will legal battles continue to lag behind innovation? The coming years will determine whether copyright frameworks can truly adapt to the age of AI.
Authored by: Ms. Mandvi Singh
She is an IP and cross-border disputes professional with expertise in patents, litigation, and global legal frameworks. With experience managing IP portfolios in the defense sector and working with top firms, she excels in patent strategy, dispute resolution, and tech law. Currently pursuing an LLM at Université Toulouse 1 Capitole, she blends legal acumen with business insights to drive innovation and compliance.
Leave a Reply