Written by: Eric Goldman
Artificial intelligence (AI) is now able to produce music, images, words, and evevideos; things that were only created by the human imagination earlier. Due to this new technology, copyright issues are being raised: who owns copyright over AI-generated work, how do we protect it, and what are the implications for human creators?
To investigate those questions, the U.S. Copyright Office brought together a panel of top economists and conducted a roundtable in early 2024. What they discovered is documented in a February 2025 report titled “Identifying the Economic Implications of Artificial Intelligence for Copyright Policy,” the second in a contemplated series of four reports. In this article, we will discuss what this second report presented.
What’s the Purpose of the Report?
This report does not call for new laws. Rather, it calls for people to consider how AI impacts copyright from an economic point of view.
The economic balance presented by copyright law was acknowledged in the U.S. Constitution, which gives creators the exclusive right to exploit their creative output for a limited period in exchange from granting the public access to that creative output. That public access was deemed essential for the advancement of the arts and sciences, and still is. While copyright law has evolved significantly over the past 250 years, this basic economic bargain remains at its core.
The rise of AI threatens that economic bargain in unique ways, and also threatens the continued advancement of the arts and sciences. The threat takes several forms, the most prominent of which is the devaluation of creative output, which means creators can face challenges in earning a living. If we as a society generate less creative output, everyone suffers. Paradoxically, the further development of AI will also be hampered by a decrease in creative output because the materials available to train AI systems will be exhausted.
The report provides a helpful distinction. Some AI models are trained on a small, known assemblage group of material, called “bounded” models. Others are trained on huge databases that are scraped from the whole internet, which are known as “unbounded” models. Economic (as well as legal) risk is much greater with unbounded models, since the content owners typically do not know that their work is being used, and they have no say in it.
Key Issues the Report Discusses
The report covers eight key topics. The most important points are the following:
1. Should AI-Generated Content Be Protected by Copyright?
Currently, there is no copyright protection for works created wholly by AI, although there is protection for works which utilize AI in the creative process. This is because of long-standing copyright policy that only output created by humans is eligible for copyright protection. For example, photographs taken by primates are not eligible for copyright protection.
The report reviewed pro and con arguments for protecting the copyright in works created by AI. On the one hand, AI does not need the economic incentive offered by copyright protection, since AI does not tire, can produce things affordably and speedily, and does not require payment or inspiration. In addition, granting copyright to works created by AI might damage human artists by curbing the market for their products and displacing creators.
On the other hand, some works created by AI can be of authentic value and contribute to the advancement of the arts and sciences. In addition, the lower cost of AI-generated content may empower users priced out of the market for human-generated creative output, from businesses looking to create advertising campaigns to singers looking to record songs. Thus, there is no clear-cut answer as to whether society as a whole would benefit from the granting of copyright protection to AI-generated content.
2. Is AI Copying Other People’s Work?
The issue of copyright infringement by AI systems generally arises in two ways. First, AI-generated output may be substantially similar to works protected by copyright; such similarity is often a necessary requirement for successful infringement claims. This is a straightforward analysis, and the economic impact of protecting creators from unauthorized copying is well established. (A closely related consideration is that of originality – specifically, a work is not even copyrightable if it is not original).
The second scenario is less simple, as the creation of substantially similar output is not involved. AI systems are trained by accessing large amounts of data, much of which is protected by copyright. If the copyright owner did not consent to the use of their work product in the training of a given AI system, the exclusive rights granted under copyright protection may be infringed. A legitimate question exists as to whether such training is actually copying, or if it falls under the general rubric of fair use.
In both situations, there are economic implications in who should be held liable for infringement. By way of comparison, the Digital MillenniumMillenium Copyright Act, or DMCA, protects internet service providers from liability when infringers use the service. While this allowed internet service providers to flourish, it meant that people whose work was infringed had to go after individual users rather than one large potential defendant (the internet service provider – who was acting, in effect, like a distributor), making litigation more costly. A similar outcome can be expected if the owners of AI systems are given protection from liability, meaning that plaintiffs would have to pursue the individual users of those AI systems –, even in litigation where the crux of the claim is was not substantial similarity but the use of copyright protected works in training the AI system in question.
3. Using Someone’s Name, Voice, or Face
Name, image and likeness, or “NIL,” rights, are not widely protected by copyright (some – but not most – states do have limited protections for image or likeness, which has resulted on occasion in successful NIL infringement suits, e.g., in the case of sound-alikes of famous individuals, Midler v. Ford Motor Co. [1988]). However, NIL rights are considered to be copyright- adjacent and are protected by a patchwork of state laws. The rise of AI-generated digital copies, so-called “deepfakes,” is presenting some unique economic questions when unauthorized uses are made of living and deceased people’s faces, likenesses and voices.
Not all uses of deepfakes are made for economic gain, as in the case of revenge porn. However, if there is no monetary value given to NIL rights, these non-economic infringements would have no remedy. For this reason, among others, there is legislation contemplated at the federal level that would give economic value to NIL rights. In many cases, there is an attempt to balance the economic implications of such infringing activity against First Amendment rights. Some proposed legislation would also distinguish between claimants who exploit NIL rights for commercial gain and claimants who do not.
4. How Training AI Affects Human Creators
The report also includes then turned to a discussion of the economic impact that training AI systems has on creators. As noted above, AI systems are trained on enormous databases, which are filled with copyrighted content. If such content is used without permission, attribution or compensation, it can lower the incentive for individuals to produce original content.
The marketplace and legislators are exploring possible solutions, but no clear consensus has emerged. One possibility is the creation of a statutory licensing system, similar to that currently used for certain music publishing rights. Another solution involves marking digital content such that trainers of AI systems are put on notice that the marked digital content is off-limits; some systems seek to potentially degrade would actually cause harm to AI systems that use marked works without authorization.
The concern is that, if some balance is not reached which meets the needs of all parties, creators might not feel at ease publishing their content online, thus undermining the public access to protected works at the heart of the copyright system.
5. Should AI Developers Get Free Access to Copyrighted Work?
The economic impact of the requirement for large amounts of data involved in training AI systems presents novel economic challenges. The larger the pool of training material, the more efficient the AI systems become. That could lead to better tools and services for everyone. But it also involves an increasingly large number of creators who are not being paid when their content is being utilized for training purposes. Compensating copyright holders for the use of their materials in training AI systems may end up serving as a competitive disadvantage to smaller AI proprietors. Any proposed solution to the problems presented by training AI systems must balance the public good of better AI systems against the need for fair compensation for creators.
6. Developers’ Access to Training Data
The article notes that a baseline should be established as to the effectiveness of AI systems trained only on public domain data. After that, there must be an analysis of the variable value of copyright protected data used to train AI systems, with the value being determined by the extent to which specific data improves output. Determining that value is part of the process of determining fair compensation to rights owners.
As of right now, there is no universal and efficient way of negotiating for access to data, as some data is controlled by large libraries, some data is controlled by publishers, and some data is controlled by individual rights owners. One idea is that creators may be able to choose whether they want to opt in or out of training AI. But that would need new, complex and costly systems for registering and managing rights, as well as confirming whether the opt-out election has been honored. Those systems might be unduly burdensome for rights holders and fundamentally change the current nature of copyright protections.
7. Controlling the Use of Copyrighted Materials in Training AI
The report also includesturns to an as-yet unasked question: is controlling the use of copyrighted material in training AI systems practical; and if it is, how can copyright policy objectives best be served? It is assumed that most existing AI systems have been trained on copyrighted material without permission, and no solution to that challenge has emerged from the marketplace. At one end of the spectrum, unrestricted and uncompensated use of copyrighted material could be allowed in training AI, in an approach that mimics a fair use argument. The other end of the spectrum is a prohibition against the unauthorized use of copyrighted materials in training AI systems, with a statutory damages scheme being incorporated into the approach. A middle ground approach is to grant amnesty for past infringements but to require permission moving forward. Another middle ground would adopt a compulsory licensing scheme like that used in the music publishing context.
8. Fairness and Bias in AI Policy
If AI is predominantly trained on the work of affluent or famous artists, it may ignore work from marginalized groups. And if human creators are not paid enough due to AI, fewer can afford to live in creative professions. That might reduce diversity in art, music, journalism, and so on; it might also undermine the ability of AI systems to advance and improve in a manner that accurately reflects society as a whole.
The Bigger Picture?
The report reminds us why copyright exists in the first place: to solve a market problem. Without copyright, creators will not produce anything new because they can’t make a living. But if copyright protection is too strong, important knowledge and culture may be withheld from the broader public.
AI complicates this issue. It reduces the price of content generation and can actually make human work unnecessary for some types of creative labor. Therefore, we must ask ourselves: how do we stimulate new work and protect creators, as well as support beneficial technology, in this new age of AI?
Final Thoughts
AI is already transforming the way content is created and shared. The economic stakes are high for artists, companies, and the public. The Copyright Office’s This report gives us a thoughtful way to start thinking about what is fair, what works, and what sort of copyright system we require for a world driven by AI.
Don’t hesitate to contact us to discuss how we can help you navigate the legal concerns of content usage or creation in light of burgeoning yet already world-changing AI systems.
