Artificial Intelligence

This post was originally published on COMMUNIA by Teresa Nobre and Leander Nielbock Last week we saw the first draft of the long-anticipated own-initiative report on copyright and generative artificial intelligence authored by Axel Voss for the JURI Committee (download as a PDF file). The report, which marks the third entry of the Committee’s recent push on the topic after a workshop and the release of a study in June, fits in with the ongoing discussions around Copyright and AI at the EU-level. In his draft, MEP Voss targets the legal uncertainty and perceived unfairness around the use of protected works and other subject matter for the training of generative AI systems, strongly encouraging the Commission to address the issue as soon as possible, instead of waiting for the looming review of the Copyright Directive in 2026. A good starting point for creators The draft report starts by calling the Commission to assess whether the existing EU copyright framework addresses the competitive effects associated with the use of protected works for AI training, particularly the effects of AI-generated outputs that mimic human creativity. The rapporteur recommends that such assessment shall consider fair remuneration mechanisms (paragraph 2) and that, in the meantime, the Commission shall “immediately impose a remuneration obligation on providers of general-purpose AI models and systems in respect of the novel use of content protected by copyright” (paragraph 4). Such an obligation shall be in effect “until the reforms envisaged in this report are enacted.” However, we fail to understand how such a transitory measure could be introduced without a reform of its own. Voss’s thoughts on fair remuneration also require further elaboration, but clearly the rapporteur is solely concerned about remunerating individual creators and other rightholders (paragraph 2). Considering, however, the vast amounts of public resources that are being appropriated by AI companies for the development of AI systems, remuneration mechanisms need to channel value back to the entire information ecosystem. Expanding this recommendation beyond the narrow category of rightholders seems therefore crucial. Paragraph 10 deals with the much debated issue of transparency, calling for “full, actionable transparency and source documentation by providers and deployers of general-purpose AI models and systems”, while paragraph 11 asks for an “irrebuttable presumption of use” where the full transparency obligations have not been fully complied with. Recitals O to Q clarify that full transparency shall consist “in an itemised list identifying each copyright-protected content used for training”—an approach that does not seem proportionate, realistic or practical. At this stage, a more useful approach to copyright transparency would be to go beyond the disclosure of training data, which is already dealt with in the AI Act, and recommend the introduction of public disclosure commitments on opt-out compliance. A presumption of use—which is a reasonable demand—could still kick in based on a different set of indicators. Another set of recommendations that aims at addressing the grievances of creators are found on paragraphs 6 and 9 and include the standardization of opt-outs and the creation of a centralized register for opt-outs. These measures are very much in line with COMMUNIA’s efforts to uphold the current legal framework for AI training, which relies on creators being able to exercise and enforce their opt-out rights. Two points of concern for users At the same time that it tries to uphold the current legal framework, the draft report also calls for either the introduction of a new “dedicated exception to the exclusive rights to reproduction and extraction” or for expanding the scope of Article 4 of the DSM Directive “to explicitly encompass the training of GenAI” (paragraph 7). At first glance, this recommendation may appear innocuous—redundant even, given that the AI Act already assumes that such legal provision extends to AI model providers. However, the draft report does not simply intend to clarify the current EU legal framework. On the contrary, the report claims that the training of generative AI systems is “currently not covered” by the existing TDM exceptions. This challenges the interpretation provided for in the AI Act and by multiple statements by the Commission and opens the door for discussions around the legality of current training practices, with all the consequences this entails, including for scientific research. The second point of concern for users is paragraph 13, which calls for measures to counter copyright infringement “through the production of GenAI outputs.” Throughout the stakeholder consultations on the EU AI Code of Practice, COMMUNIA was very vocal about the risks this category of measures could entail for private uses, protected speech and other fundamental freedoms. We strongly opposed the introduction of system-level measures to block output similarity, since those would effectively require the use of output filters without safeguarding users rights. We also highlighted that model-level measures targeting copyright-related overfitting could have the effect of preventing the lawful development of models supporting substantial legitimate uses of protected works. As this report evolves, it is crucial to keep this in mind and to ensure that any copyright compliance measures targeting AI outputs are accompanied by relevant safeguards that protect the rights of users of AI systems. A win for the Public Domain One of the last recommendations in the draft report concerns the legal status of AI-generated outputs. Paragraph 12 suggests that “AI-generated content should remain ineligible for copyright protection, and that the public domain status of such works be clearly determined.” While some AI-assisted expressions can qualify as copyright-protected works under EU law —most importantly when there’s sufficient human control over the output—many will not meet the standards for copyright protection. However, these outputs can still potentially be protected by related rights, since most have no threshold for protection. This calls into question whether the related rights system is fit for purpose in the age of AI: protecting non-original AI outputs with exclusive rights regardless of any underlying creative activity and in the absence of meaningful investment is certainly inadequate. We therefore support the recommendation that their public domain status be asserted in those cases. Next steps Once the draft report is officially published and presented in JURI on

Luca Schirru Recently, a Danish Bill has been making headlines by addressing issues related to deepfake through a rather uncommon approach: copyright. As stated to The Guardian, the Danish Minister of Culture, Jakob Engel-Schmidt, explained that they “are sending an unequivocal message that everybody has the right to their own body, their own voice and their own facial features, which is apparently not how the current law is protecting people against generative AI.” According to CNN, the minister believes that the “proposed law would help protect artists, public figures, and ordinary people from digital identity theft.” Items 8, 10, and 19 of the proposal include some of the most substantive changes to the law. Among other measures, Item 8 proposes adding a new § 65(a), requiring the prior consent of performers and performing artists to digitally generate imitations of them and make these available to the public, establishing protection for a term of 50 years after their death. Item 10 introduces a new § 73(a), focusing on “realistic digitally generated imitations of a natural person’s personal, physical characteristics,” requiring prior consent from the person being imitated before such imitations can be made available to the public. This exclusive right would also last for 50 years after the death of the imitated person and would not apply to uses such as caricature, satire, parody, pastiche, criticism, or similar purposes. It could be argued that this approach is uncommon because several countries, including those in the European Union, already have laws regulating personality rights and, more specifically, personal data. Copyright is known for regulating the use of creative expressions of the human mind, not the image, voice, or likeness of a person when considered individually, i.e., outside the context of an artistic performance. According to CNN “Engel-Schmidt says he has secured cross-party support for the bill, and he believes it will be passed this fall.”  A machine-translated version of the Proposal is below:  Notes:

Luca Schirru ‘Boring’ is not a word that can be used to describe the past few days for those interested in litigation involving copyright issues in the development and use of Generative AI systems. Two major cases saw significant updates, issuing orders that addressed one of the main questions raised in these lawsuits: is the use of copyrighted materials to train Generative AI systems fair use? This blog post aims to briefly describe each case’s key points related to fair use and to highlight what was left unresolved, including all the ‘what if’ scenarios that were hinted at but not decided upon Bartz, Graeber & Johnson v. Anthropic Judge William Alsup’s order on fair use addressed not only the different copies of copyrighted material made for training generative AI systems but also uses related to Anthropic’s practice of keeping copies as a “permanent, general-purpose resource”. It also distinguished between legally purchased copies and millions of pirated copies retained by Anthropic, applying a different fair use analysis to each category. Regarding the overall analysis of fair use for copyrighted works used to train Anthropic’s Generative AI system, Judge Alsup found that the use “was exceedingly transformative and was a fair use.” Among the four factors, only the second factor weighed against using copyrighted works to train the GenAI system. Concerning the digitization of legally purchased books, it was also considered fair use not because of the purpose of training AI systems, but for a much simpler reason:  “because all Anthropic did was replace the print copies it had purchased for its central library with more convenient space-saving and searchable digital copies for its central library — without adding new copies, creating new works, or redistributing existing copies”. For this specific use, of the four factors, only factor two weighed against fair use, while factor four remained neutral. On the other hand, Judge Alsup clearly stated that using pirated copies to create the “general-purpose library” was not fair use, even if some copies might be used to train LLMs. All factors weighed against it. Specifically, Judge Alsup noted: “it denies summary judgment for Anthropic that the pirated library copies must be treated as training copies. We will have a trial on the pirated copies used to create Anthropic’s central library and the resulting damages, actual or statutory (including for willfulness).” Kadrey v. Meta At the very beginning of the order, Judge Vince Chhabria clarified that the case questions whether using copyrighted material to train generative AI models without permission or remuneration is illegal and affirmed that: “although the devil is in the details, in most cases the answer will likely be yes. What copyright law cares about, above all else, is preserving the incentive for human beings to create artistic and scientific works. Therefore, it is generally illegal to copy protected works without permission. And the doctrine of “fair use,” which provides a defense to certain claims of copyright infringement, typically doesn’t apply to copying that will significantly diminish the ability of copyright holders to make money from their works (thus significantly diminishing the incentive to create in the future).” Judge Chhabria explained further that  “by training generative AI models with copyrighted works, companies are creating something that often will dramatically undermine the market for those works, and thus dramatically undermine the incentive for human beings to create things the old-fashioned way.” According to him, this would primarily affect not classic works or renowned authors but rather the market for the “typical human-created romance or spy novel,” which could be substantially diminished by similar AI-created works.  However, all these points were framed as “this Court’s general understanding of generative AI models and their capabilities”, with Judge Chhabria emphasizing that “Courts can’t decide cases based on general understandings. They must decide cases based on the evidence presented by the parties.”  Despite this general understanding that “copying the protected works, however transformative, involves the creation of a product with the ability to severely harm the market for the works being copied, and thus severely undermine the incentive for human beings to create“, Judge Chhabria found two of the plaintiffs’ three market harm theories “clear losers,” and the third, a “potentially winning” argument, underdeveloped: “First, the plaintiff might claim that the model will regurgitate their works (or outputs that are substantially similar), thereby allowing users to access those works or substitutes for them for free via the model. Second, the plaintiff might point to the market for licensing their works for AI training and contend that unauthorized copying for training harms that market (or precludes the development of that market). Third, the plaintiff might argue that, even if the model can’t regurgitate their own works or generate substantially similar ones, it can generate works that are similar enough (in subject matter or genre) that they will compete with the originals and thereby indirectly substitute for them. In this case, the first two arguments fail. The third argument is far more promising, but the plaintiffs’ presentation is so weak that it does not move the needle, or even raise a dispute of fact sufficient to defeat summary judgment.“ In the overall analysis of the four factors, only the second factor weighed against Meta. Summary judgment was granted to Meta regarding the claim of copyright infringement from using plaintiffs’ books for AI training. Nevertheless, Judge Chhabria clarified that “this ruling does not stand for the proposition that Meta’s use of copyrighted materials to train its language models is lawful. It stands only for the proposition that these plaintiffs made the wrong arguments and failed to develop a record in support of the right one.” The use of pirated copies was also addressed in Kadrey v. Meta. In this case, “there is no dispute that Meta torrented LibGen and Anna’s Archive […].” According to Judge Chhabria, while downloading from shadow libraries wouldn’t automatically win the plaintiffs’ case, it was relevant for the fair use analysis, especially regarding “bad faith” and whether the downloads benefited or perpetuated unlawful

Lokesh Vyas and Yogesh Badwal This post was originally published on Spicy IP. In the previous part, we examined whether the opt-out mechanism, as claimed in Gen-AI litigations, constitutes a prohibited formality for the “enjoyment and exercise” of authors’ rights under Article 5(2) of the Berne Convention. And we argued no. In this post, we address the second question: Can opting out be permitted as an exception under the three-step test outlined in Article 9(2)? If you haven’t seen the previous post, some context is helpful. (Or, you can skip this part) As we mentioned in the last post, “Many generative AI models are trained on vast datasets (which can also be copyrighted works) scraped from the internet, often without the explicit consent of content creators, raising legal, ethical, and normative questions. To address this, some AI developers have created and claimed “opt-out mechanisms,” allowing copyright holders or creators to ask that their works not be used in training (e.g., OpenAI’s Policy FAQs).  Opt out under the Copyright Exception A  question arises here: What are the other ways opt-out mechanisms can be justified if the states want to make a mechanism like that? One may say that opt-outs can be valid under the Berne Convention if an exception (e.g., an AI training exception with an inbuilt opt-out possibility) passes the three-step test. And this way, opt-outs can be regarded as a legitimate limit on holders’ exclusive rights. For reference, the three-step test was created in the 1967 revision conference, later followed in Article 13 of TRIPS and Article 10 of WCT. The test creates a room for the nations to make certain exceptions and limitations. Article 9(2) authorises the member countries “to permit the reproduction” of copyright works in 1.) “certain special cases, provided that such reproduction 2.) does not conflict with a normal exploitation of the work and 3.) does not unreasonably prejudice the legitimate interests of the author”.  Although we don’t delve into the test, how opting out can be a part of an exception can be understood from an example. For instance, as Ginsburg exemplifies, if a country states that authors lose their translation rights unless they explicitly reserve or opt out of them, it would violate Article 5(2) because such rights under Berne must apply automatically, without formalities. This actually happened with Turkey in 1931, whose application for membership was rejected due to the condition of deposit for translation rights in its domestic law. (See Ricketson and Ginsburg’s commentary, paragraph 17.18.)  But if an exception (like allowing radio retransmissions in bars) already complies with Berne’s provisions and applies equally to all authors, then letting authors opt out of that exception would give them more rights than Berne requires. And this should be permissible.  Notably, introducing an exception, such as for AI training, must first pass the three-step test. Opt out can be built therein. However, remember that every exception presupposes a prima facie infringement. Within that frame, the opt-out offers the author a chance not to lose. Thus, it creates an inadvertent expansion of her rights beyond the convention.  Additionally, opt-out can fare well with the three-step test due to the factor of “equitable remuneration to authors.” As Gompel notes in his piece, “…‘opt out’ eases compliance with the three-step test because it mitigates some of the adverse effects of the proposed copyright exception. That is, it enables authors to retain exclusivity by opting out of the compensation scheme.”  Another question also exists: Did Berne contain particular provisions that directly allowed an opt-out arrangement? Well, the answer is Yes. Does opting out equal the right to reserve under Article 10bis? Not really. Setting aside the debate over formality and the three-step test, the Berne Convention contains an opt-out-style provision, albeit limited, where authors must explicitly reserve their rights to avoid specific uses of their work. Relevant here is Article 10bis of the Convention, which allows member countries to create exceptions for the reproduction of works published in newspapers on, among other topics, current economic, political, or religious issues. However, it also allows the authors to ‘expressly reserve’ their work from reproduction. Indian Copyright Act, 1957 also contains a similar provision in Section 52(1)(m). Interestingly, the right to reserve exploitation has been part of the Berne Convention since its earliest draft. It first appeared in Article 7 alongside the provision on formalities, which was numbered Article 2 in the draft. Article 7 became Article 9(2) in 1908, when formalities were prohibited and the no-formality rule entered the Berne Convention.  This historical pairing raises a strong presumption: opting out of a specific mode of exploitation cannot automatically be deemed a prohibited formality. Ginsburg confirms this, citing the 1908 Berlin Conference, which clarified that the reservation/opt-out clause (then Article 9(2)) was not considered a formality. But can this special setting (created in Article 10bis(1)) be used to open the door for general opt-out AI exception measures by countries? We doubt it. As the negotiation history of the 1967 revision conference suggests, Article 10bis(1) is a lex specialis, i.e., a narrow and specific exception (See page 1134 of Negotiations, Vol. II). This means that it may derogate from the general no-formalities rule, but it cannot serve as a model for broader declaratory measures.  Conclusion The upshot is that opt-outs may be de facto formalities. However, not all formalities are prohibited under the Berne Convention. The convention enables countries to make some formalities on “the extent of protection.” Three key points emerge from this discussion: One, opting out may not be a formality that prevents the enjoyment and exercise of rights, as Gompel and Sentfeln confirm, and Ginsburg argues otherwise. Two, it can be a part of an AI training exception if such an exception can pass the three-step test. When applying this test, opting out would support the factor of equitable remuneration. Three, Article 10(bis) on the right to reserve cannot be read expansively. While it can be used to justify the three-step test as Sentfleben does, it might not be extended generally. Okay. That’s it from our end. À bientôt’ Primary Sources:-

Lokesh Vyas and Yogesh Badwal This post was originally published on Spicy IP. Bonjour, Lately, we’ve been cogitating on this curious concept called the “opt-out”, which has been cropping up with increasing frequency in generative AI litigation, including in India. The EU and the UK are taking the idea seriously and considering giving it statutory teeth. On the surface, it is sold as a middle path, a small price to pay for “balance” in the system. However, at least prima facie, it seems like a legal absurdity that fractures its modern foundational logic, where authors receive default copyright without any conditions. The opt-out model, the argument goes, reintroduces formality through the back door, a de facto formality of sorts. This shifts the burden onto authors and rights holders to actively monitor or manage their works to avoid unintended inclusion in the AI training. There have been questions about whether such an opt-out scheme is compatible with the Berne Convention, which prohibits the same under Article 5(2), e.g., here, here, and here.  Given the complex nature of this issue and the fact that many such discussions happen behind paywalls, making them inaccessible to the public, we thought it would be beneficial to share our ideas on this topic and invite further reflection. This two-part post mainly focuses on the legality of opting out without addressing its implementability and applicability, which raises several questions (e.g., as discussed recently in Martin Sentfleben’s post). In short, we probe whether opt-outs violate the Berne Convention—the first international copyright law treaty binding on all members of the TRIPS and WCT.  We answer it through two questions and discuss each one separately. First, is opt-out a prohibited formality for the “enjoyment and exercise” of authors’ rights under Article 5(2) of the Berne Convention? Two, can it be permitted as an exception under the three-step test under Article 9(2)? We answer the first question in the negative and the second in the positive. Additionally, we also examine whether Berne already has a provision that can allow this without looking at the details.  This post addresses the first question. What Makes Opts Outs So Amusing – The Flip? Many generative AI models are trained on vast datasets, which can also include copyrighted works scraped from the internet without the explicit consent of content creators, raising legal, ethical, and normative concerns. To address this, some AI developers have created and claimed “opt-out mechanisms,” allowing copyright holders or creators to ask that their works not be used in training (e.g., OpenAI’s Policy FAQs).  Herein lies the catch: it requires authors and copyright holders to explicitly say “No” to training by adding a robots.txt tag to their website with specific directives that disallow web crawlers from accessing their content. (E.g., see this OpenFuture’s guide here) Thus, instead of creators being protected by default, they are supposed to opt out to prevent exploitation. One could say that this flips the logic of copyright on its head–from a presumption of protection to a presumption of permission. But that’s not so simple.  Notably, opting out is not a novel argument. In fact, it can be dated back at least to the 1960s in the Nordic countries’ model of “Extended Collective Licensing” (ECL), which mandates collective licensing while preserving the author’s right to opt out. Other notable academic literature on opt-out can be found here, here, here, and here, dating back over two decades. Swaraj also covered this issue a decade ago. In particular, we must acknowledge the scholarship of Jane Ginsburg, Martin Sentfleben, and Stef van Gompel, who have significantly influenced our thinking on the topic. Two Key Questions: Opt out as a Formality and opt out under a permitted Exception Formality Argument first.  Here, the argument goes that the opt-out is a prohibited formality under Article 5(2) and should not be allowed. However, we doubt it. Let’s parse the provision first. Which states: “(2) The enjoyment and the exercise of these rights shall not be subject to any formality; such enjoyment and such exercise shall be independent of the existence of protection in the country of origin of the work. Consequently, apart from the provisions of this Convention, the extent of protection, as well as the means of redress afforded to the author to protect his rights, shall be governed exclusively by the laws of the country where protection is claimed.” (Authors’ emphasis) For context, the provision pertains to “Rights Guaranteed outside the Country of Origin” for both national and foreign authors. And the question of no-formality pertains particularly to foreign authors. In other words, by removing formality requirements in the country where protection is claimed, the provision enabled authors to automatically receive protection without needing to satisfy foreign formalities. This matters because while countries can impose conditions on their own nationals, it’s generally assumed that they will not treat their own authors worse than foreign ones. The post follows this presumption: if a country cannot burden foreign authors, it’s unlikely to impose stricter terms on its own people. Although the removal of formalities had been discussed in the international copyright law context as early as the 1858 Brussels Conference, an important event in the development of international copyright law, it was not implemented until 1908. This change addressed practical difficulties, including identifying the “country of origin” when a work was published in multiple countries, and the challenges courts faced in enforcing rights without formalities. (See International Bureau’s Monthly Magazine, January 1910) Tellingly, while a country can make formalities for its people, it cannot do so for foreign authors. It’s generally assumed that a country would not obligate its authors more than it does to foreign authors. Textual Tensions of Article 5(2) While the phrase “any formality” in the first line of the provision might suggest that all kinds of formalities—including de facto ones like opt-out mechanisms—are prohibited, that is arguably not the case. We say this because the provision is divided into two parts, and the prohibition on formalities applies only to the first part, which is germane to enjoying and exercising rights. The second part of the provision, beginning with “Consequently”, gives leeway to the states wherein they can make formalities regarding the ‘extent of protection’ and

Luca Schirru and Sean Flynn At the beginning of May, the report “Copyright and Artificial Intelligence. Part 3: Generative AI Training” was released, sparking a wide range of debates due to its content and the political issues surrounding its release. In this short contribution, we aim to briefly introduce the report and touch on some of the key content and political issues currently being discussed. SCOPE AND STRUCTURE OF THE REPORT The first thing that stands out about the report appears right on its first page: “pre-publication version”: a label reported as unusual and potentially unprecedented. The 113-page document addresses one of the most controversial issues at the intersection of copyright and Generative AI: the use of protected content to train Generative AI systems.  While most sources focus on fair use, the report also includes sections on “technical background,” “infringement,” and “licensing for AI training”, all of which are a “must read,” especially for those just joining the discussion and feeling overwhelmed by the hundreds of thousands of articles, blogs, books, and other resources available on the topic. The report attempts to summarize some of the main issues in both the legal and technical fields. The approach taken by the USCO is sometimes described as “favorable to copyright owners” or as “a mixed bag”, receiving both praise and criticism on multiple fronts, as we will illustrate below. POLITICAL CONTEXT AND CONTROVERSIES The timing of the report While it may be early to determine the precise reasons behind the (unusual) release of a pre-publication version, several explanations have been speculated, though none have been confirmed. The report states that its early release was made “in response to congressional inquiries and expressions of interest from stakeholders.” However, questions have been raised that may relate to concerns about potential restrictions under Trump Administration, which is arguably aligned with positions favorable to big technology companies, as well as fear that the report could be buried in the event of the dismissal of the Register of Copyright, or the potential influence on ongoing legal cases. Regarding the latter, there have been concerns about the timing of the report and how it could interfere with the outcomes, especially the fair use analysis, of ongoing lawsuits. As noted, “it could put a thumb on the scale for how the courts will resolve these cases,” without giving the parties an opportunity to address any potential gaps in the report, which could have a significant impact on other GenAI cases. Leadership changes and copyright policy While the timing of the notice of dismissal of Shira Perlmutter (Register of Copyrights at the time the report was drafted) and the release of the report could give rise to the inference that the report was the sole reason for her dismissal, other events may have influenced the decision as well. The day before the release of the (pre-publication version of the) report, the Librarian of Congress, Carla Hayden, who had appointed Shira Perlmutter, was dismissed. Therefore, concerns about additional leadership changes may also have played a role in the decision to release the pre-publication version. An extra layer of complexity arises when one considers that Perlmutter’s position was one appointed and overseen by the legislative branch. The argument that the report may have contributed to the dismissal has often been linked to an alleged alignment between the Trump Administration’s position and that of big tech companies. This connection can be inferred from Rep. Joe Morelle’s statement, reported by POLITICO, claiming it is “no coincidence [Trump] acted less than a day after [Perlmutter] refused to rubber-stamp Elon Musk’s efforts to mine troves of copyrighted works to train AI models.” Finally, as reported by Authors Alliance, on April 30, “American Accountability Foundation urges President Trump to fire ‘deep state’ librarians, targeting Carla Hayden and Shira Perlmutter,” based on the claim that Hayden was supporting Biden policies, particularly in the areas of intellectual property and transgender rights. FAIR USE AT THE HEART OF THE DEBATE While the report addresses multiple issues, both legal and technical, the most debated (and anticipated) topics are those related to whether the use of protected content to train Generative AI systems qualifies as fair use. The fair use chapter is the longest in the report, comprising nearly half of its content. It includes a factor-by-factor analysis applied to different scenarios, with the USCO identifying the first and fourth factors as taking on particular prominence in the analysis. In the section titled “weighing the factors,” the Office states the following: “As generative AI involves a spectrum of uses and impacts, it is not possible to prejudge litigation outcomes. The Office expects that some uses of copyrighted works for generative AI training will qualify as fair use, and some will not. On one end of the spectrum, uses for purposes of noncommercial research or analysis that do not enable portions of the works to be reproduced in the outputs are likely to be fair. On the other end, the copying of expressive works from pirate sources in order to generate unrestricted content that competes in the marketplace, when licensing is reasonably available, is unlikely to qualify as fair use. Many uses, however, will fall somewhere in between.” (p.74) While there has been some agreement with certain parts of the report, such as the acknowledgment that litigation outcomes cannot be prejudged, and the view that “research and academic uses should be favored under the fair use analysis”, one of the most criticized aspects is the interpretation of the fourth factor in the fair use analysis, in which the Report concludes that original works created by AI that are not substantially similar to works used in the training may nonetheless result in “market dilution” that should weigh against a fair use analysis. According to USCO’s report: “While we acknowledge this is uncharted territory, in the Office’s view, the fourth factor should not be read so narrowly. The statute on its face encompasses any “effect” upon the potential market.373 The speed and scale at

Luca Schirru About the Report In February 2025, the U.S. Copyright Office released the report “Identifying the Economic Implications of Artificial Intelligence for Copyright Policy: Context and Direction for Economic Research”, edited by USCO’s chief economist, Brent Lutes. The report was produced after months of research, interactions among scholars and technical experts, and the outcomes of a roundtable event. By identifying the most pressing economic issues related to copyright and artificial intelligence (AI), the roundtable “aimed to provide a structured and rigorous framework for considering economic evidence so that the broader economic research community can effectively answer specific questions and identify optimal policy choices.” Considering the length of the report and the variety and complexity of the issues it addresses, we will split our analysis into two separate blog posts: one focusing on the output phase and the other on the input phase. Following the structure of the report, we will begin with the output-related topics: “Copyrightability of AI-Generated Works and Demand Displacement” and “Copyright Infringement by AI Output”, as these are most directly connected to copyright. For this reason, we will not summarize the section on “Commercial Exploitation of Name, Image, and Likeness”, and instead recommend that readers refer directly to the report for details on that topic.  Copyrightability of AI-Generated Works and Demand Displacement This chapter, whose principal contributors are Imke Reimers and Joel Waldfogel, proposes the following question: “how the emergence of generative AI technology affects the optimal provision of copyright protection?” When discussing whether AI-generated works should be copyrighted, it connects to whether they cause a net positive value, and that there would also be the need “to be weighed against the value of human-generated works displaced by the technology”. (p.10) The substitution effect is also considered, not only in cases where AI-generated works substitute human-generated ones, but also when AI-generated works are verbatim or near-verbatim reproductions of pre-existing human-generated content. Similarly, some of these near-verbatim reproductions may decrease the value of the related human work when, for example, they provide misinformation. Such a decrease in value may also reduce interest in human-generated works. On the other hand, and from an economic perspective, the report also suggests that “all of its uses would supplant revenue for human creators. Some uses will reduce deadweight loss, replacing it with consumer surplus by allowing for additional consumption that otherwise would not occur”. (p.10) One of the effects that may be seen in the long run relates to the fact that human experimentation leads to more radical stylistic innovation and experimentation, while it is not clear “whether AI-generated output can ever engage in the same sort of experimentation and innovation as humans”. (p.11) While the report acknowledges that there is a possibility that AI may reduce production costs and be a tool to promote creativity, increase productivity, and enhance quality, it warns about the risk of less experimentation, crowding out “more risky and costly experimental creations that sometimes lead to valuable innovation”. (p.11) Displacing human creators may even be harmful to the development of Gen AI, as these models are trained with human-generated works, according to the report. A first conclusion that may be drawn from this section is that further research, including empirical research, needs to be carried out to better understand issues like the value created and displacement caused by GenAI, the decrease in the value of human-generated works, the “degree to which the fixed cost recovery problem exists for AI-generated works” (p.12), and “the demand curve and cost function for creative works”. (p.14)  When it comes to offering copyright protection to these AI-generated outputs, the report suggests that it would incentivize their production and affect human output in both positive and negative ways. However, it also recalls that this may not be optimal, as “copyright inherently limits public access to existing works and thus produces a social cost”. (p.12) The report also notes that production costs may differ between human-generated and AI-generated works, and that “copyright protection only serves its economic objective if the social value of the former outweighs that of the latter. If the fixed production costs of AI-generated works are sufficiently low, the additional incentives of copyright are not necessary for reaching optimal production levels, thus, offering copyright protection would be suboptimal”. (p.12) Copyright Infringement by AI Output As previously mentioned, the report does not delve into legal issues, focusing instead on economic analysis. In the chapter primarily contributed by Joshua Gans, the author offers considerations from an economic perspective on defining the “optimal scope of what output is infringing,” noting that “copyright protection from infringement should balance the incentives to produce and the ability to consume creative works.” The author begins by explaining one of the structural dynamics of copyright, where “the mechanism used for incentivizing the production of new works (exclusive rights pertaining to the usage of a work) also limits consumers’ access to existing works”, and that the “broadest possible scope of protection could also effectively hinder new creative output for fear of liability”. (p.16) It argues that an important step in the analysis is to identify the “optimal level of market power that we wish to confer to rightsholders in the context of competing AI-generated works”, assuming this level to be the same as that used in infringement disputes involving human-generated works. The chapter proposes considering multiple, but not all, factors that may impact the balance mentioned above, and reflects on how this would be different in cases involving AI (pp.16-17) Several factors may affect the market power of the rightsholders, including but not limited to the threshold for infringement (the higher the threshold, the lower the power) and the requirements to demonstrate that the copy was infringing. On the latter, it is also argued that in the cases concerning AI-generated work, “access [to the allegedly infringed work] may be harder to dispute”. (p.17)  According to the study, these factors may be helpful to understand if a rightsholder may or may not exercise its market power, but the potential