Barely a day passes without some controversy surrounding the topic of Generative AI. If it’s not fake videos of naked women or a provocative South Park skit, it’s a court case involving a student who created a song with lyrics dangerously similar to those of Taylor Swift.
One thing is clear: Gen AI is truly game-changing. We can analyze and summarize massive quantities of research, mock up an image for a game or ad, even if we are artistically lacking, and knock out a quick Python script that exceeds our current skill level as programmers.
However, all these tools have emerged from a largely US-based Wild West culture of AI development. In order to train the next generation of large language models (LLMs), engineers have ridden roughshod over copyright protections, sucking in seemingly every available source of digitized information. Current copyright court cases spend their time litigating whether someone’s work was used as the input for training, or if a similarity to it has been generated at the output.
Here in Europe, where the creative industries contribute €354 billion to the economy [1], people are naturally worried that the hard work invested in their art and literature is being used for a previously unforeseen purpose, with no resultant financial recompense. Not only that, but they’re also hearing that Gen AI will offer a fantastic future for the creative industries; however, because they’re humans, they’re not going to be part of it.
Naturally, the EU’s legislators have been taking a look, developing some regulatory frameworks. But will this strangle what is still a nascent technology, or will it reset the balance between the tech bros and the creative industry’s output that they rely upon?
The CDSM Directive and the AI Act
Copyright came about during another technological revolution – the invention of the printing press. Suddenly, written works could be mass-produced rather than waiting for your local monastery to handwrite a copy. Since then, copyright has moved with the times, accommodating changes in media output and technology. And this is where we find ourselves once more, with computers performing data processing far beyond that of any human-based endeavor.
Unlike the US, the EU has no “fair use” doctrine, but there are copyright exceptions for situations such as research and private study, criticism and parody, and archival and heritage [2]. The Copyright in the Digital Single Market (CDSM) Directive provides a legal framework for Text and Data Mining (TDM), recognizing that data is required to enable advances in the world of AI.
Specifically, Article 3 makes provision for TDM by scientific research organizations. Article 4 is for everyone else, including commercial AI developers. Here, there is an exception for rights holders to opt out, even if the published information is available online, enabling them to instead make the content available through a licensing agreement [3].
These provisions are underscored by the EU AI Act, the world’s first comprehensive law for this technology, built around copyright, transparency, data governance, and ethical considerations. It requires those developing AI models to summarize the training data used, enabling copyright holders to enforce their rights.
From scraping to licensing
The naysayers frame these regulations as restrictive, placing unnecessary hurdles in the place of progress, and throttling a potentially massive EU industry sector that is already on the back foot. Others, like ourselves at ape factory, would prefer to describe them as providing a much-needed framework, or scaffolding, to build a trustworthy AI market.
Here’s an example of what’s at stake. Hayao Miyazaki of Studio Ghibli has seen his artistic style replicated by users of Generative AI, including ChatGPT CEO Sam Altman [4]. Such mass creation of images in his unique style damages the value of his existing body of work, resulting in people now questioning the authenticity of any image they see in that style. Is this the world we really want to live in?
The Article 4 opt-out clause resolves issues like this, forcing tech companies to enter licensing agreements with such artists if they wish to do so. The resultant AI platforms would then be able to offer access to the selected artistic or literary styles they have in their portfolio. In many ways, this is no different from media companies bidding to show Premier League football.
In fact, this process has already started, with Wiley, John Hopkins University Press, and Taylor & Francis/Informa agreeing content licenses with a range of tech companies [5]. The EU also expects the role of Content Management Organisations (CMOs) to grow, handling licensing and collecting royalties on behalf of smaller creators.
Why Gen AI developers only benefit
The “free speech” and anti-regulation drumbeaters who have gotten this far will, I suspect, remain unconvinced. And I doubt there is much to say that could convince them otherwise – that’s just how society rolls currently. But European investors and established industry players really should be breathing a sigh of relief.
The AI Act provides legal certainty following a litany of high-profile legal copyright cases, some of which are still pending. In Kneschke v. LAION (Germany), the non-profit LAION benefited from an Article 3 CDSM exemption when providing text-image datasets, as they fell under TDM for scientific research. Other, similar cases are still working their way through the courts.
In the U.S., the case of Bartz v. Anthropic PBC resulted in a partial summary. Some of the data used by Anthropic came from books they had purchased and scanned, which was considered fair use. However, judgment was refused for claims based on training using pirated copies of books that had been downloaded, and the case will proceed to trial. There are also numerous ongoing class-action lawsuits by visual artists and media organizations in the U.S.
Even with this small sample of litigation, it’s clear that the EU approach has its benefits. Gen AI businesses here have a clear framework within which to operate. No doubt there will be judicial tests at its boundaries, but the general consensus is that rights holders will be better served by regulation than litigation. In the U.S., neither the ‘AI Foundation Model Transparency Act of 2023’ nor the 'Generative AI Copyright Disclosure Act of 2024’ were passed into law. This leaves them with litigation as the only path for those artists who have deep pockets – and those without will simply miss out.
Responsible and prosperous AI for the EU
The following conclusion is difficult to write: EU lawmakers seem to have made some good decisions. Yes, generally speaking, we could do with fewer rules and regulations and more government efficiency in Europe. But the AI Act and CDSM actually seem to deliver some much-needed certainty around what has felt like the tech bros’ ‘smash and grab’ job in the creative industries. We’re no legal scholars, but we do seem to have a well-calibrated framework that protects creators, provides legal certainty for Gen AI innovators, and offers enough flexibility to foster new markets.
As discussed in the EUIPO’s (European Union Intellectual Property Office) report on the topic [6], far from strangling AI, the EU seems to have shaped a digital future within which technical progress and creative industries can thrive, together. It remains true to European values, grounded in fairness, while enabling us to innovate. “This is not just a study,” says João Negrão, the EUIPO Executive Director, “it is a call to coordinated, forward-looking action.” We hope it is.
*** BREAKING NEWS *** It turns out that dealing with copyright for AI training post-use costs $1.5 billion (around $3,000 per book). Thanks to Anthropic for the clarification [7].
-------
[1] https://www.eif.org/InvestEU/guarantee_products/ie-ccs-market-study.pdf
[3] 2025_GenAI_from_copyright_perspective_executive_brief_en.pdf
[4] https://www.cullenllp.com/blog/ai-and-artistic-style-imitation-emerging-copyright-implications/
[5] https://sr.ithaka.org/our-work/generative-ai-licensing-agreement-tracker/
