Five Problems with Foreign Government AI Copyright Proposals
Artificial Intelligence continues to evolve at a breakneck pace, and governments around the world are examining their laws and racing to propose policies and regulations intent on encouraging AI innovation. Australia, the European Union, France, Hong Kong, India, Singapore, South Korea, Thailand, and the United Kingdom have all considered or adopted legal reforms to address AI copyright issues.
Copyright law is one of the central battlegrounds in the AI regulatory landscape. However, in the rush to legislate on the issue, many AI copyright related proposals by foreign governments have been shaped by incorrect assumptions that do not align with the realities of copyright law, international legal frameworks, or existing AI copyright licensing markets—thus undermining copyright law and its robust incentive structure that supports human creativity.
There are five things foreign governments tend to get wrong when proposing AI copyright-related policies and regulations. These include:
1. Copyright Isn’t a Barrier to AI Development
A common assumption underlying these problematic AI copyright proposals put forth by foreign governments is the idea that copyright law is an obstacle to innovation—that copyright is “blocking” or slowing down AI development, and thus government intervention is necessary to remove these so-called “barriers” in the global race for AI dominance. In reality, copyright has always been fundamental to fueling technological industries. Without music, film and television shows, videos, photographs, news articles, there would be a real lack of social media platforms, streaming platforms, and other technologies that we enjoy and use on a daily basis. Generative AI technologies are the same. Without copyrighted works, AI technologies would not have the generative capabilities they do today. Through providing a legal framework for how creative works may be accessed, copied, and reused, copyright is an important economic incentive and foundation for continued human creativity and development that inevitably also fuels AI development.
Strong copyright protections incentivize the creation of high-value works, which in turn form the basis of the high-quality datasets that AI systems rely on. When it comes to AI systems, as the saying goes, “quality in, quality out.” That is, AI outputs are only as reliable as the data on which they are trained. Research has shown that when models are trained on lower-quality or synthetic data, particularly when they continuously train on their own outputs, performance degrades over time, resulting in a loss of diversity in responses, bias amplification, and confidently incorrect responses. In other words, increasing the quantity of synthetic data cannot substitute for the quality of human-generated content. This reinforces the significance of copyrighted works, not only as inputs for AI systems, but also as a necessary foundation for maintaining the performance and reliability of generative AI technologies.
Treating copyright as a barrier to innovation risks promulgating reforms that would harm the creative community and inevitably undermine AI development. A more sustainable approach is one that recognizes that incentivizing and rewarding those who create high-quality works is key to development of high-quality AI systems. Therefore, foreign governments should take an approach to encourage creations and investments in copyrighted works by upholding and strengthening copyright laws, not weakening and undermining them, which will be necessary to promoting the long-term success of both the creative and technological sectors.
2. There is Mounting Evidence of an AI Licensing Market
In proposing or enacting an exception for AI use of copyrighted works, foreign governments often ignore the existence of a robust AI licensing market. In fact, licensing agreements between rightsholders and AI developers were in place even before the exponential rise of generative AI technologies around 2022. As generative AI models became technologically viable and commercially profitable, the market for AI copyright licensing grew exponentially. Since the end of 2022, rightsholders offered and reached agreements with AI developers and companies on thousands of licensing arrangements that permit the use of copyrighted works in a wide range of generative AI use scenarios and contexts. These AI licensing agreements span a range of industries—including publishing, news media, music, and stock media and visual licensing services—demonstrating that licensing is not only possible but also increasingly common.
Policymakers of foreign governments should therefore be cautious about accepting claims that licensing markets are nonexistent or that licensing is impossible in the AI context. If anything, the exponential growth in the number and kind of licensing deals indicates that clearer legal intent to protect copyrights in the context of generative AI technologies may encourage further licensing. But overly broad copyright exceptions risk prematurely undermining markets that are naturally emerging out of necessity and industry buy-in.
3. International Agreements Limit the Type and Scope of Permissible AI Copyright Exceptions
A recurring flaw in many of the proposed AI copyright exceptions proposed by foreign governments is that many of these exceptions run afoul of major international obligations such as the three-step test of the Berne Convention for the Protection of Literary and Artistic Works (Berne Convention) and the World Trade Organization Agreement on Trade-Related Aspects of Intellectual Property (TRIPS Agreement).
Under Article 9(2) of the Berne Convention and Article 13 of the TRIPS Agreement, member countries can permit exceptions to the right of reproduction of copyrighted works in (1) certain special cases, (2) provided that such reproduction does not conflict with a normal exploitation of the work, and (3) does not unreasonably prejudice the legitimate interests of the author. Broad AI copyright exceptions proposed by foreign governments fail these three steps. AI copyright exceptions fail this three-step test because:
- First, such an exception would fail the “special case” requirement. Exceptions and limitations which authorize a broad scope of use, such as one which permits all unauthorized text-and-data-mining uses conducted for any purpose related to AI development or for AI “training,” is not a “special case.” Such exceptions are simply general authorizations for mass reproduction of protected works without any further limitations or qualifications, including by specifying AI use cases or addressing commerciality. In the AI context, the concept of “use” encompasses a wide range of distinctive activities—copying, ingestion, analysis, storage, and downstream dissemination and use—each of which may implicate different exclusive rights and policy considerations. Framing an exception around such an ambiguous term collapses these distinct actions into a single category, further removing the exception from the specificity requirement under the Berne Convention. Accordingly, any exception would have to be narrowly tailored to very specific uses for very specific purposes to meet the “special case” requirement of the Berne Convention.
- Second, exceptions tend to fail the second step because they directly conflict with existing and normal exploitation of the works. As previously mentioned, there is an existing and growing market for licensing copyrighted works for all sorts of AI applications and uses. These licenses are often the product of direct negotiations between rightsholders and AI developers but also include collective agreements negotiated on behalf of rightsholders.
According to some estimates, collective licensing systems already generate more than $16 billion annually in global licensing revenue, demonstrating that markets for authorized uses of copyrighted works are both active and economically significant to individual creators and artists. At the same time, a rapidly growing market for AI training data, already valued in the low billions and projected to exceed $11 billion within the next decade, indicates that licensing models are actively expanding to accommodate new technological uses while rewarding copyright owners, which can encompass collective licensing models. When copyright laws are upheld and enforced, this incentivizes robust free market exchanges that unlocks the greatest potentials from these licensing markets that benefit both the creative and tech economies. Sweeping exceptions to copyright laws undermine these licensing markets, as AI developers would be able to use copyrighted works without paying the customary price, thus interfering with the existing and normal exploitation of the works.
- Third, AI copyright exceptions that have been proposed by foreign governments risk unreasonably prejudicing the legitimate interests of the authors. Copyright is built on an incentive structure that enables creators to control and benefit from the use and exercise of the copyrights to their works. Allowing AI developers to copy troves of copyrighted works without authorization from the author or copyright owner undermines their fundamental ability to exercise control over the copyright to their works and the underlying economic incentive structures arising from copyright protections. This loss of control and economic leverage is precisely the kind of prejudice the third step is intended to prevent.
Authors’ legitimate interests are further prejudiced where AI outputs of AI models involuntarily trained on works become a substitute for and compete against the original copyrighted works that were copied. The AI model’s output does not need to be identical to the original copyrighted work to be a plausible substitute for the work. If the training process enables a model to generate outputs that compete with, replicate, or displace demand for the original works, a broad AI copyright exception is no longer simply facilitating innovation, but is actively shifting value away from creators and allowing the generated works to undermine the same market as the original trained-on content. In this case, the competition of the generated work actively prejudices the author’s legitimate interests in commercial exploitation of their work.
4. Opt-Out Regimes Don’t Work
Opt-out regimes are often presented as a compromise solution that purportedly aids AI development while offering rightsholders a measure of control over whether their works may be used for AI training. However, this approach subverts the foundational rule of copyright: a copyright owner has the right to choose whether to authorize the use of their work. Under an opt-out model, the default is effectively reversed, and creators are compelled to affirmatively prevent unauthorized use of their works or otherwise have their works used to train AI models. As a general matter, copyright owners should not be required to take proactive steps to maintain the rights they inherently possess under copyright law, particularly where the uses involve large-scale reproductions for commercial gain.
This burden-shifting is not only a policy concern but also raises serious issues with international copyright norms. Opt-out requirements could effectively function as a copyright formality, risking another Berne Convention violation. Article 5 of the Berne Convention states:
(1) Authors shall enjoy, in respect of works for which they are protected under this Convention . . . the rights which their respective laws do now or may hereafter grant to their nationals, as well as rights specially granted by this Convention.
(2) The enjoyment and the exercise of these rights shall not be subject to any formality …
This prohibition on formalities reflects a core principle of the Berne Convention, which is that authors should not be required to take specific affirmative steps to secure copyright. “Formalities” typically include things like copyright notice and term renewals, each of which requires some action on the part of the rightsholder as a condition of protection. An opt-out requirement functions essentially in the same way by requiring a rightsholder to complete a procedural step to prevent reproduction of their work, whereas a rightsholder who fails to act forfeits control over uses that should otherwise require authorization.
Further, Article 5(2) specifically prohibits formalities affecting both the “enjoyment” and the “exercise” of rights protected under the Convention. An opt-out regime burdens the exercise of the copyright by requiring rightsholders to engage in continuous vigilance and affirmative objection to use across innumerable platforms and jurisdictions. This is precisely the type of complication the Berne Convention was designed to prevent.
The same concern applies to the three-step test discussed in the previous section. The mere existence of an opt-out regime does not address any of the harms arising from an otherwise overbroad exception under the three-step test including conflicts with the normal exploitation of the work, the displacement of existing and emerging licensing markets, and unreasonable prejudice to the legitimate interests of rightsholders. Shifting the burden of the action from users to rightsholders does not eliminate the existence of prejudice. It simply redistributes the costs of compliance and places them on the party arguably least able to bear them. This is exacerbated by how the design of many opt-out processes may work against users, with one study finding that dozens of tech and AI companies use deceptive and confusing design patterns in their consumer opt-out processes. On the other hand, an opt-in approach more appropriately places the onus on AI developers as the copyright users to identify which works they want to use and to secure the necessary permissions to use them. Not only does this avoid issues with burden-shifting and formalities, but arguably AI developers are in a better position to determine the need for specific works than a rightsholder would be to ensure they have properly opted out of all possible training datasets.
In addition, the practical effects of opting out are often negligible for a variety of reasons. First, once a model has been trained on a dataset, it may be impossible to remove. Second, despite a creator opting out, the work may still be included in the dataset through other means, such as the inclusion of an unauthorized copy in a dataset scraped from a licensee of the copyright owner or from a third-party platform where a pirated copy has been posted. Finally, tools designed to facilitate opt-out also tend to be limited, as many of the current tools were not designed to address scraping and ingestion for AI training in the first place. As a result, opt-out regimes offer a minimal ability at best for rightsholders to control their works in the AI context, but are instead more of a symbolic, empty gesture, as they create the perception of choice without offering creators any meaningful control or enforceability.
5. Lawful Access Requirements are Ineffective and Misleading
Some proposed AI training copyright exception frameworks have relied on “lawful access” requirements, suggesting that AI developers should be permitted to use copyrighted works for training as long as the works were lawfully accessed. However, this approach is misleading because it conflates lawful access with lawful use. The fact that a work can be and/or was accessed legally, whether through purchase, subscription, or public availability, does not mean that the work may be copied, reproduced, or included in an AI training dataset. Simply because a subscriber may listen to a song on a streaming service does not grant them permission to make copies of the song or play that song for commercial purposes. In such cases, although the access of the work was lawful, any subsequent use of the work would likely violate the terms of service for the subscription.
Lawful access requirements also fail to address the reality of how large-scale AI training often occurs in practice. Typically, many copyrighted works, including pirated or unauthorized copies, are uploaded to publicly accessible websites and may be downloaded or scraped indiscriminately. Under a lawful access framework, AI developers could argue that scraping these works to train AI models is permissible because it was publicly accessible, despite the underlying upload having been unauthorized. A lawful access requirement could end up driving creators and publishers to further restrict their works through paywalls or closed platforms to prevent scraping, which would ultimately harm consumers by limiting the availability of high-quality, lawful content online. Thus, lawful access proposals risk blurring the legal distinction between access and use of works, to the detriment of rightsholders and consumers alike.
Conclusion
Global policymakers are analyzing, examining, and proposing statutory changes with the goal of supercharging AI development or attracting AI investments and companies to their countries. However, many of the AI copyright-related proposals to achieve these goals are built on flawed misconceptions. This blog covered five things foreign governments get wrong when proposing AI copyright changes, such as treating copyright law as a barrier to innovation, undermining fundamental principles of copyright in ways that harm creators and copyright owners, and disregarding substantial evidence by ignoring overwhelming evidence of a robust AI copyright licensing market—but there are other issues as well, which you can read about from the Copyright Alliance’s submissions in response to foreign government proposals about AI and copyright.
A durable, long-term approach will recognize that copyright and technological advancement are not incompatible. A system that respects copyright will not only protect human creativity and culture but also lead to more sustainable growth and development of AI technologies.
If you aren’t already a member of the Copyright Alliance, you can join today by completing our Individual Creator Members membership form! Members gain access to monthly newsletters, educational webinars, and so much more — all for free!