We are pleased to announce that we have submitted a comment to the Copyright Office in response to their recent notice of inquiry regarding how copyright law interacts with generative AI. In our comment, we shared our views on copyright and generative AI (which you can read about here) and the stories we heard from authors about how they are using generative AI to support their creative labors, research, and the mundane but important tasks being involved with being a working author. The Office received over 10,000 comments in response to its NOI, showing the high level of interest in how copyright regulates AI-generated works and training data for generative AI. We hope the Office will appreciate our perspective as it considers policy interventions to address copyright issues involved in the use of generative AI by creators. You can read our full comment here, or at the bottom of this post. 

You can hear more about our comment, and about contributions from other commenters, at the Berkeley Center for Law and Technology virtual roundtable on Monday, November 13th, where Authors Alliance senior staff attorney Rachel Brooke will be a panelist. The event is free and open to the public, and you can sign up here. 

Background

Since the Copyright Office issued an opinion letter on copyright in a graphic novel containing AI-generated images back in February, the debate about copyright and generative AI has grown to a near fever pitch. Authors Alliance has been engaged in these issues since the decision letter was released: we exist to support authors who want to leverage the tools available in the digital age to see their creations reach broad audiences and create innovative new works, and we see generative AI systems as one such tool that can support authors and authorship. We participated in the Copyright Office’s listening session on copyright issues in AI-generated textual works this spring, and were eager to further weigh in as the Copyright Office wades through the thorny issues involved. 

In late August, the Copyright Office issued a notice of inquiry, asking stakeholders to weigh in on a series of questions about copyright policy and generative AI. These were broken down into general questions, questions about training AI models, questions about transparency and recordkeeping, and various issues related to AI outputs—copyrightability, infringement, and labeling and identification. 

Our Comment

Our comment was devoted in large part to sharing the ways that authors are using generative AI systems and tools to support their creative labors and research. We heard from authors that used generative AI systems for ideation, late stage editing, and generating text. We also learned that authors are using generative AI systems in ways we wouldn’t have anticipated—like creating books of prompts for other authors to use as inputs for generative AI systems. Generative AI has helped authors who don’t publish with conventional publishers create marketing copy and even generate book covers (despite the common adage, these are pretty important for attracting readers). We also heard from researchers using generative AI for literature reviews as well as to make their writing process more efficient so they can focus on doing the work of researching and innovating. Generative AI also has the potential to lower barriers to entry for scientific researchers who are not native English speakers, but want to make contributions to scientific fields in which literature tends to be written in English. 

We also spent some time explaining our views on why the use of copyrighted materials in training datasets for AI models constitutes fair use and how fair use analysis applies when copyrighted materials are included in training datasets. The use of creative works in training datasets is a transformative one with a different purpose than the works themselves—regardless of whether the institutions that develop and deploy them are commercial or nonprofit. And it’s highly unlikely that a generative AI system could harm the markets for the works in the training sets for the underlying models: a generative AI system is not a substitute for a book a reader is interested in reading, for example. We also explained that the market harm consideration (factor four in fair use analysis) should consider the effect of the use (using training data on AI models) on the market for the specific work in question (i.e., in an infringement action, the work that is alleged to have been infringed), and not the market for that author’s other works, similar works, or anything else.

Our comment also argued that new copyright legislation on AI—either to codify copyright’s human authorship requirement and explain how it applies to AI-generated content or to address other issues related to copyright and generative AI—is not warranted. AI systems, AI models, and the ways creators use them are still evolving. Copyright law is already highly flexible, having adapted to new technologies that weren’t anticipated when the copyright legislation itself was enacted. And legislating around nascent technologies can result in laws that are eventually ill-suited to deal with unexpected challenges that new technologies bring about (recall that the DMCA, which has faced a lot of criticism as a statute intended to regulate copyright online, was passed in 1998). We instead suggest that the Office stick with a “wait and see” approach as generative AI and how we use it continue to develop rather than recommending legislation to Congress. 

Next, we explained why a licensing system for AI works in training data is neither desirable nor practicable. Because we consider the use of copyrighted works in training data to be a fair use, licenses are not necessary in the first place. We also explained the host of problems that either a compulsory licensing regime or a collective licensing scheme would bring about. The large size of datasets for training AI models make it difficult to envision systematically seeking licenses for each and every copyrighted work in the training dataset, and the “orphan works problem” means that a majority of rightsholders might not be able to be found. It’s also not clear who would administer licensing under a licensing regime, and we could not think of any appropriate party that exists or is likely to emerge. The Office’s past failed investigations into possible collective rights management organizations (or CMOs) only underscore this point. 

Finally, we echoed our support for the substantial similarity test as a way to handle generative AI outputs that look very similar to existing copyrighted works. The substantial similarity test has been around for decades and has been applied across the country in a variety of contexts. It seems to us to be a good way to approach the rare cases in which generative AI outputs are strikingly similar to copyrighted works (so-called “memorization”) such that a rightsholder might sue for infringement. 

What’s Next?

The same day we submitted our comment, the Biden Administration released an executive order on “Safe, Secure, and Trustworthy Artificial Intelligence,” directing federal agencies to take a variety of measures to ensure that the use of generative AI is not harmful to innovation, privacy, labor, and more. Then on Wednesday, representatives from a coalition of countries (including the U.S.) signed “The Bletchley Declaration” following an AI Safety Summit in the U.K., warning of the dangers of generative AI and pledging to work together to find solutions. All of this is to say that how public policy should regulate generative AI, and whether and how the law needs to change to accommodate it, is a live issue that continues to evolve every day. Dozens of lawsuits are pending about the interaction between copyright and the use of generative AI systems, and as these cases move through the courts, judges will have their opportunity to weigh in. As ever, we will keep our readers and members appraised in any new legal developments around copyright and generative AI. 

COLC-2023-0006-8976_attachment_1

Authors Alliance is currently at work on a submission to the Copyright Office regarding our views on generative AI (which you can read about here). If you’re an author who has used generative AI in your research or writing, we’d love to hear from you! Please reach out to Rachel Brooke, Authors Alliance Senior Staff Attorney, at rachel@authorsalliance.org.

Last week, the Court of Appeals for the D.C. Circuit released its opinion in the American Society for Testing and Medical Materials v. Public.Resource.org (“ASTM v. PRO”), an important fair use case that has been percolating in the D.C. Circuit for the past few years. Authors Alliance filed an amicus brief in the case in support of Public Resource, along with the Library Futures Institute, the EveryLibrary Institute, and Public Knowledge. The case is about public access to the law and the role of fair use in safeguarding that access, but it also has big implications for the ever-evolving doctrine of fair use. In general, we applaud the decision, which found for Public Resource, affirming the importance of access to the law and the important role that the fair use doctrine plays within copyright law. In today’s post, we summarize the case and offer our thoughts about what it might mean for fair use going forward, particularly regarding cases that impact our members and their interests.

Background

The case concerns standard-developing organizations and public access to the standards they produce. These organizations set standards and best practices for “particular industries, products, or problems,” including fire prevention and medical testing, among others. These standards are often incorporated into laws and regulations that govern these industries by various federal, state, and local lawmaking bodies. Government agencies incorporate these standards into law “by reference” when they refer to them in a given regulation, without reproducing the standards verbatim. For example, a federal regulation governing shipyard operators requires them to “select, maintain, and test portable fire extinguishers” in accordance with a particular National Fire Protection Association standard, but that regulation does not reproduce the standard itself. 

Public.Resource.org, a nonprofit organization that disseminates legal materials by posting them publicly online, posted on its website “hundreds of incorporated standards—including standards produced and copyrighted by the plaintiffs.” Then, in 2013, the standard-developing organizations sued for copyright infringement. Public Resource defended its posting of the standards as a fair use, but the lower court disagreed, requiring Public Resource to take the posted standards at issue down. After appeals, further fact development and multiple hearings at both the district court and appellate court level, the district court ultimately found Public Resource’s posting of the standards which were incorporated into law to be fair use. The standard-developing organizations appealed to the appeals court, which released its decision on September 12th.

Our Amicus Brief

In our amicus brief, we argued that “when a law-making body incorporates a standard by reference into legally-binding rule or regulation, the contents of the whole of that publication must be freely and fully accessible by the public.” Public access to the law is crucial for an informed citizenry and well-functioning democracy, which is why more conventional legal materials—like statutes, regulations, court cases, and agency rulemakings—have long been freely available to the public, online or otherwise. This principle ought to extend to legal standards that are incorporated by reference into law, despite the fact that private organizations create these standards, because incorporation by reference essentially gives them the force of law. We emphasized the potential harm to researchers and librarians were public access to standards incorporated by reference into law restricted. 

In fact, our brief argues that these standards should not be afforded copyright protection at all. Allowing private organizations to claim copyright in what is effectively the law does not serve the core purpose of copyright—to incentivize new creation for the benefit of the public. Materials authored by the federal government are automatically a part of the public domain, which also supports the important principle that no one can own the law—an idea which is enshrined in our Constitution and court cases dating to the 19th century. Due process—a Constitutional principle requiring the legal rights of all persons to be respected—mandates this kind of access, and it is often painted as one that is “beyond question.” While the standard-setting organizations have online “reading rooms” where the public can access the standards in question, this requires users to register, provide personal information, and agree to lengthy terms of service. As we explain in our brief, this is not sufficient for the free public access that the law requires.

The Decision

In its decision, the court determined that Public Resource’s posting of the standards that were incorporated by reference into law was a fair use, holding that three out of the four fair use factors favored a finding of fair use. While the court did not hold that the standards incorporated by reference into law were free from copyright protection, it did affirm the legal and policy justifications for free public access to the law. 

The first fair use factor, the purpose and character of the use, weighed in favor of Public Resource. On this point, the court emphasized that “Public Resource’s use is for nonprofit, educational purposes.” The question of whether a use is commercial can impact the way a court views this factor, as can the degree to which a court finds the use to be “transformative.” The court similarly found that Public Resource’s use was transformative, in that it was new and different from the purpose of the works themselves. Unlike the purposes of the original standards developed by the organizations—to promulgate best practices for industries and problems in the interest of industries and consumers—Public Resource’s purpose was to share with the public “only what the law is, not what industry groups may regard as current best practices.” The court summarized: “Public Resource’s message (‘this is the law’) is very different from the plaintiffs’ message (‘these are current best practices for the engineering of buildings and products’).” 

The second fair use factor directs courts to consider the nature of the copyrighted work—in this case, the standards that were incorporated by reference into law. The court found that this factor strongly favored a finding of fair use. The further a work from the “core of intended copyright protection,” i.e., the more creative it is, the more this factor favors fair use. In other words, because the standards at issue were highly factual in nature, rather than creative (like fiction writing), the second factor weighed in favor of fair use. 

The third fair use factor considers the amount and substantiality of the portion of the original work that was used, asking whether the portion of the work that was used is reasonable in light of the purpose of the secondary user’s use. The court found that this factor also weighed in favor of fair use. The various standards promulgated by the standard-setting organizations tended to be much longer in their entirety than the portions that were incorporated by reference into law. Public Resource only posted the portions of these standards that were incorporated into law, which was of course reasonably in light of its purpose of educating the public about what the law is. 

The fourth fair use factor considers the effect of the use on the market for the copyrighted works, and the court found that this factor was, on balance, neutral, and “[did] not significantly tip the balance one way or the other.” The standard setting organizations argued that their customers—industry members that needed to understand best practices—would fail to pay for the standards if they could obtain them for free from Public Resource. The court pointed out that only the standards incorporated into law were at issue, and the most up-to-date standards relied on by these industries were not necessarily incorporated into law. Moreover, the standard-setting organizations could not actually produce any evidence of market harm, despite the fact that Public Resource had been posting them online for approximately 15 years. The court also indicated that the public benefit of sharing this information with the public had to be balanced against any potential market harm. But because there was a possibility that Public Resource’s online posting could have lowered demand for the standards, the court found that this factor was neutral.

Impact on the Fair Use Doctrine

It remains to be seen how this case will impact the fair use doctrine and fair use decisions going forward, but it seems quite likely that this new judicial precedent might make a difference in future fair use decisions.

First, the contours of factor one—the purpose and character of the use—are very much a live issue following the recent decision in Warhol Foundation v. Goldsmith. In that case (in which we also submitted an amicus brief, supporting the Warhol Foundation’s fair use argument), the Supreme Court emphasized the fact that Warhol’s use was commercial in finding the use not to be fair. It seemed to emphasize commerciality over “transformativeness,” a longstanding aspect of factor one analysis (though that court found the use to not be transformative). The court in ASTM v. PRO certainly discussed commerciality as part of factor one, emphasizing Public Resource’s nonprofit status. But regarding the question of transformativeness, the court also gave a lengthy and eloquent summary of the different purposes of the two uses, indicating that transformativeness is still an important inquiry, and is not necessarily secondary to commerciality.

The weight of commerciality in factor one analysis can make a big difference in the outcome of cases, and it is an issue many have been watching with the dearth of copyright lawsuits concerning the use of copyrighted works to train generative AI models. This is because while there is a strong argument that the use of training data for these models is highly transformative, it is also true that the companies behind many of the models—like OpenAI, Midjourney, and Stability AI—are commercial in nature, and monetize their programs in different ways. The recent ASTM v. PRO decision could affect how courts weigh the commerciality of these companies’ uses of copyrighted training data against the extent to which the uses are transformative, potentially tipping the scale towards fair use in the upcoming copyright lawsuits about generative AI and training data. 

Second, the question of market harm in factor four can be a complicated one, and this case may provide some guidance for courts going forward. This issue was animated in the recent decision in Hachette Books v. Internet Archive—the case about whether controlled digital lending is a fair use, which we have been covering and involved in for years now, notably as an amicus in support of the Internet Archive. In the Hachette decision, the judge found that factor four weighed in favor of the publishers without direct evidence of financial harm, based on the idea that CDL scans could be substitutes for licensed ebooks. But in ASTM v. PRO, the court was skeptical that an allegation of potential market harm, without actual evidence, was sufficiently convincing. Since Hachette has been appealed and will soon be before the Second Circuit, we are hopeful that ASTM v. PRO will be a useful precedent for those judges. Extending the logic of ASTM v. PRO, it may be that the publishers will need to demonstrate market harm with tangible evidence (such as concrete evidence of lost sales) in that case in order to prevail on factor four.

Authors Alliance readers will surely have noticed that we have been writing a lot about generative AI and copyright lately. Since the Copyright Office issued its decision letter on copyright registration in a graphic novel that included AI-generated images a few months back, many in the copyright community and beyond have struggled with the open questions around generative AI and copyright.

The Copyright Office has launched an initiative to study generative AI and copyright, and today issued a notice of inquiry to solicit input on the issues involved. The Senate Judiciary Committee has also held multiple hearings on IP rights in AI-generated works, including one last month focused on copyright. And of course there are numerous lawsuits pending over its legality, based on theories ranging from copyright infringement to to privacy to defamation. It’s also clear that there is little agreement about a one-size-fits-all rule for AI-generated works that applies across industries. 

At Authors Alliance, we care deeply about access to knowledge because it supports free inquiry and learning, and we are enthusiastic about ways that generative AI can meaningfully further those ideals. In addition to all the mundane but important efficiency gains generative AI can assist with, we’ve already seen authors incorporate generative AI into their creative processes to produce new works. We’ve also seen researchers incorporate these tools to help make new discoveries. There are some clear concerns about how generative AI tools, for example, can make it easier to engage in fraud and deception, as well as perpetuating disinformation. There have been many calls for legal regulation of generative AI technologies in recent months, and we wanted to share our views on the copyright questions generative AI poses, recognizing that this is a still-evolving set of questions.  

Copyright and AI

Copyright is at its core an economic regulation meant to provide incentives for creators to produce and disseminate new expressive works. Ultimately, its goal is to benefit the public by promoting the “progress of science,” as the U.S. Constitution puts it. Because of this, we think new technology should typically be judged by what it accomplishes with respect to those goals, and not by the incidental mechanical or technological means that it uses to achieve its ends. 

Within that context, we see generative AI as raising three separate and distinct legal questions. The first and perhaps most contentious is whether fair use should permit use of copyrighted works as training data for generative AI models. The second is how to treat generative AI outputs that are substantially similar to existing copyrighted works used as inputs for training data—in other words, how to navigate claims that generative AI outputs infringe copyright in existing works. The third question is whether copyright protection should apply to new outputs created by generative AI systems. It is important to consider these questions separately, and avoid the temptation to collapse them into a single inquiry, as different copyright principles are involved. In our view, existing law and precedent give us good answers to all three questions, though we know those answers may be unpalatable to different segments of a variety of content industries. 

Training Data and Fair Use

The first area of difficulty concerns the input stage of generative AI. Is the use of training data which includes copyrighted works a fair use, or does it infringe on a copyright owner’s exclusive rights in her work? The generative AI models used by companies like OpenAI, Stability AI, and Stable Diffusion are based on massive sets of training data. Much of the controversy around intellectual property and generative AI concerns the fact that these companies often do not seek permission from rights holders before training their models on works controlled by these rights holders (although some companies, like Adobe, are building generative AI models based on their own stock images, openly-licensed images, and public domain content). Furthermore, due to the size of the data sets and nature of their collection (often obtained via scraping websites), the companies that deploy these models do not make clear what works make up the training data. This question is one that is controversial and highly debated in the context of written works, images, and songs. Some creators and creator communities in these areas have made calls for “consent, credit, and compensation” when their works are included in training data. The obstacle to that point of view is, if the use of training data is a fair use, none of this is required, at least not by copyright.  

We believe that the use of copyrighted works as training data for generative AI tools should generally be considered fair use. We base this view on our reading of numerous fair use precedents including Google Books and HathiTrust cases as well others such as iParadigms. These and other cases support the idea that fair use allows for copying for non-expressive uses—copying done as an “intermediate step” in producing non-infringing content, such as by extracting non-expressive content such as patterns, facts, and data in or about the work. The notion that non-expressive (also called “non-consumptive”) uses do not infringe copyrights is based in large part on a foundational principle in copyright law: copyright protection does not extend to facts or ideas. If it did, copyright law would run the risk of limiting free expression and inhibiting the progress of knowledge rather than furthering it. Using in-copyright works to create a tool or model with a new and different purpose from the works themselves, which does not compete with those works in any meaningful way, is a prototypical fair use. Like the Google Books project (as well as text data mining), generative AI models use data (like copyrighted works) to produce information about the works they ingest, including abstractions and metadata, rather than replicating expressive text. 

In addition, fair use of copyrighted works as training data for generative AI has several practical implications for the public utility of these tools. For example, without it, AI could be trained on only “safe materials,” like public domain works or materials specifically authorized for such use. Models already contain certain filters—often excluding hateful content or pornography as part of its training set. However, a more general limit on copyrighted content—virtually all creative content published in the last one hundred years—would tend to amplify bias and the views of an unrepresentative set of creators. 

Generative AI Outputs and Copyright Infringement

The feature that most distinguishes generative AI from technology in copyright cases that preceded it, such as Google Books and HathiTrust, is that generative AI not only ingests copyrighted works for the purpose of extracting data for analysis or search functionality, but for using this extracted data to produce new content. Can content produced by a generative AI tool infringe on existing copyrights?

Some have argued that the use of training data in this context is not a fair use, and is not truly a “non-expressive use” because generative AI tools produce new works based on data from originals and because these new works could in theory serve as market competitors for works they are trained on. While it is a fair point that generative AI is markedly different from those earlier technologies because of these outputs, the point also conflates the question of inputs and outputs. In our view, e using copyrighted works as inputs to develop a generative AI tool is generally not infringement, but this does not mean that the tool’s outputs can’t infringe existing copyrights. 

We believe that while inputs as training data is largely justifiable as fair use, it is entirely possible that certain outputs may cross the line into infringement. In some cases, a generative AI tool can fall into the trap of memorizing inputs such that it produces outputs that are essentially identical to a given input. While evidence to date indicates that memorization is rare, it does exist. 

So how should copyright law address outputs that are essentially memorized copies of inputs? We think the law already has the tools it needs to address this. Where fair use does not apply, copyright’s “substantial similarity” doctrine is equipped to handle the question of whether a given output is similar enough to an input to be infringing. The substantial similarity doctrine is appropriately focused on protection of creative expression while also providing room for creative new uses that draw on unprotectable facts or ideas. Substantial similarity is nothing new: it has been a part of copyright infringement analysis for decades, and is used by federal courts across the country. And it may well be that standards, such as a set of  “Best Practices for Copyright Safety for Generative AI” proposed by law professor Matthew Sag, will become an important measure of assessing whether companies offering generative AI have done enough to guard against the risk of their tools producing infringing outputs.

Copyright Protection of AI Outputs

A third major question is, what exactly is the copyright status of the outputs of generative AI programs: are they protected by copyright at all, and if so, who owns those copyrights? Under the Copyright Office’s recent registration guidance, the answer seems to be that there is no copyright protection in the outputs. This does not sit well with some generative AI companies or many creators who rely on generative AI programs in their own creative work. 

We generally agree with the Copyright Office’s recent guidance concerning the copyright status of AI-generated works, and believe that they are unprotected by copyright. This is based on the simple but enduring “human authorship” requirement in copyright law, which dates back to the late 19th century. In order to be protected by copyright, a work must be the product of a human author and contain a modicum of human creativity. Purely mechanical processes that occur without meaningful human creative input cannot generate copyrightable works. The Office has categorized generative AI models as this kind of mechanical tool: the output responds to the human prompt, but the human making the prompt does not have sufficient control over how the model works to make them an “author” of the output for the purposes of copyright law. The district court for D.C. recently issued a decision agreeing with this take in Thaler v. Perlmutter, a case that challenged the human authorship requirement in the context of generative AI. 

It’s interesting to note here that in the Copyright Office listening session on text-based works, participants nearly universally agreed that outputs should not be protected by copyright, agreeing with the Copyright Office’s guidance. Yet the other listening sessions had more of a diversity of views. In particular, the participants in the listening sessions on audiovisual works and sound recordings were concerned about this issue. In industries like the music and film industries, where earlier iterations of generative AI tools have long been popular (or are even industry norms), the prospect of being denied copyright protection in songs or films, simply due to the tools used, can understandably be terrifying for creators who want to make a profit from their works. On this front, we’re sympathetic. Creators who rely on their copyrights to defend and monetize their works should be permitted to use generative AI as a creative tool without losing that protection. While we believe that the human authorship requirement is sound, it would be helpful to have more clarity on the status of works that incorporate generative AI content. How much additional human creativity is needed to render an AI-generated work a work of human authorship, and how much can a creator use a generative AI tool as part of their creative process without foregoing copyright protection in the work they produce? The Copyright Office seems to be grappling with these questions as well and seeking to provide additional guidance, such as in a recent webinar with more in-depth registration guidance for creators relying on generative AI tools in their creative efforts.

Other Information Policy Issues Affecting Authors

Generative AI has generated questions in other areas of information policy beyond the copyright questions we discuss above. Fraudulent content or disinformation, the harm caused by deep fakes and soundalikes, defamation, and privacy violations are serious problems that ought to be addressed. Those uses do nothing to further learning, and actually pollute public discourse rather than enhance it. They can also cause real monetary and reputational harm to authors. 

In some cases, these issues can be addressed by information policy doctrines outside of copyright, and in others, they can be best handled by regulations or technical standards addressing development and use of generative AI models. A sound application of state laws such as defamation law, right of publicity laws, and various privacy torts could go a long way towards mitigating these harms. Some have proposed that the U.S. implement new legislation to enact a federal right of publicity. This would represent a major change in law and the details of such a proposal would be important. Right now, we are not convinced that this would serve creators better than the existing state laws governing the right of publicity. While it may take some time for courts to figure out how to adapt legal regimes outside of copyright to questions around generative AI, adapting the law to new technologies is nothing new. Other proposals call for regulations like labeling AI-generated content, which could also be reasonable as a tool to combat disinformation and fraudulent content. 

In other cases, creators’ interests could be protected through direct regulation of the development and use of generative AI models. For example, certain creators’ desire for consent, credit, and compensation when their works are included in training data sets for generative AI programs is an issue that could be perhaps addressed through regulation of AI models. As for consent, some have called for an opt-out system where creators could have their works removed from the training data, or the deployment of a “do not train” tag similar to the robots.txt “do not crawl” tag. As we explain above, under the view that training data is generally a fair use, this is not required by copyright law. But the views that using copyrighted training data without some sort of recognition of the original creator is unfair, which many hold, may support arguments for other regulatory or technical approaches that would encourage attribution and pathways for distributing new revenue streams to creators. 

Similarly, some have called for collective licensing legislation for copyrighted content used to train generative AI models, potentially as an amendment to the Copyright Act itself. We believe that this would not serve the creators it is designed to protect and we strongly oppose it. In addition to conflicting with the fundamental principles of fair use and copyright policy that have made the U.S. a leader in innovation and creativity, collective licensing at this scale would be logistically infeasible and ripe for abuse, and would tend to enrich established, mostly large rights holders while leaving out newer entrants. Similar efforts several years ago were proposed and rejected in the context of mass digitization based on similar concerns.  

Generative AI and Copyright Going Forward

What is clear is that the copyright framework for AI-generated works is still evolving, and just about everyone can agree on that. Like many individuals and organizations, our views may well shift as we learn more about the real-world impacts of generative AI on creative communities and industries. It’s likely that as these policy discussions continue to move forward and policymakers, advocacy groups, and the public alike grapple with the open questions involved, the answers to these open questions will continue to develop. Changes in generative AI technology and the models involved may also influence these conversations. Today, the Copyright Office published issued a notice of inquiry on the topic of copyright in AI-generated works. We plan to submit a comment sharing our perspective, and are eager to learn about the diversity of views on this important issue.

Last week, the District Court for the District of Columbia announced a decision in Thaler v. Perlmutter, a case challenging copyright’s human authorship requirement in the context of a work produced by a generative AI program. This case is one of many lawsuits surrounding copyright issues in generative AI, and surely will not be the last we hear about the copyrightability of AI-generated works, and how this interacts with copyright’s human authorship requirement. In today’s post, we’ll provide a quick summary of the case and offer our thoughts about what this means for authors and other creators.

Background

Back in 2018 (before the current public debate about copyright and generative AI had reached the fever pitch we now see today), Dr. Stephen Thaler applied for copyright registration in a work of visual art produced by a generative AI system he created, called the Creativity Machine. Thaler sought to register his work as a computer-generated “work-made-for-hire” since he created the machine, which “autonomously” produced the work. After a lot of back and forth with the Copyright Office, it maintained its denial of the application, explaining that the human authorship requirement in copyright law foreclosed protection for the AI-generated work, since it was not the product of a human’s creativity.

Then, Thaler then sued Shira Perlmutter, the Register of Copyrights, in the D.C. district court, asking the court to decide “whether a work autonomously generated by an AI system is copyrightable.” Judge Baryl A. Howell upheld the Copyright Office’s decision, explaining that under the plain language of the Copyright Act, “an original work of authorship” required that the author be a human “based on centuries of settled understanding” and a dictionary definition of “author.” She also cited to the U.S. Constitution’s IP clause, which similarly mentions “authors and inventors,” and over a century of Supreme Court precedent to support this principle.

Thaler’s attorney has indicated that he will be appealing the ruling to the D.C. Circuit court of appeals, and it remains to be seen whether that court will affirm the ruling. 

Implications for copyright law

The headline takeaway from this ruling is that AI generated art is not copyrightable because copyright requires human authorship, which remains a requirement in copyright law. However, the ruling is actually more nuanced and contains a few subtle points worth highlighting. 

For one, this case tested not just the human authorship requirement but also the application of the work-for-hire doctrine in the context of generative AI. On one view of the issues, if Thaler created a machine capable of creating a work that would be copyrightable were it created by a human, there is a certain appeal in framing the work as one commissioned by Thaler. On this point, the court explained that since there was no copyright in the work in the first instance based on its failure to meet the human authorship requirement, this theory also did not hold water. In other words, a work-made-for-hire requires that the “hired” creator also be a human. 

It’s important to keep in mind that Thaler was in a sense testing the reach of the limited or “thin” copyright that can be granted in compilations of AI-generated work, or AI-generated work that a human has altered, thus endowing it with at least a modicum of human creativity as copyright requires. Thaler made no changes to the image produced by his Creativity Machine, and in fact, described the process to the Copyright Office as fully autonomous rather than responding to an original prompt (as is generally the case with generative AI). Thaler was not trying to get a copyright in the work in order to monetize it for his own livelihood, but—presumably—to explore the contours of copyright in computer-generated works. In other words, the case has some philosophical underpinnings (and in fact, Thaler has said in interviews that he believes his AI inventions to be sentient, a view that many of us tend to reject). But for creators using generative AI who seek to register copyrights in order to benefit from copyright protection, things are unlikely to be quite so clear-cut. And while she found the outcome to be fairly clear cut in this case, Judge Howell observed:

“The increased attenuation of human creativity from the actual generation of the final work will prompt challenging questions regarding how much human input is necessary to qualify the user of an AI system as an ‘author’ of a generated work, the scope of the protection obtained over the resultant image, how to assess the originality of AI-generated works where the systems may have been trained on unknown pre-existing works, how copyright might best be used to incentivize creative works involving AI, and more.”

What does this all mean for authors? 

For authors who want to incorporate AI-generated text or images into their own work, the situation is a bit murkier than it was for Thaler. The case itself provides little in the way of information for human authors who use generative AI tools as part of their own creative processes. But while the Copyright Office’s registration guidance tells creators what they need to do to register their copyrights, this decision provides some insight about what will hold up in court. Courts can and do overturn agency actions in some cases (in this case, the judge could have overturned the Copyright Office’s denial of Thaler’s registration application had she found it to be “arbitrary and capricious”). So the Thaler case in many ways affirms what the Copyright Office has said so far about registrability of AI-generated works, indicating that the Office is on the right track as far as their approach to copyright in AI-generated works, at least for now. 

The Copyright Office has attempted to provide more detailed guidance on copyright in “AI-assisted” works, but a lot of confusion remains. One guideline the Office promulgated in a recent webinar on copyright registration in works containing AI-generated material is for would-be registrants to disclose the contribution of an AI system when its contribution is more than “de minimis,” i.e., when the AI-generated creation would be entitled to copyright protection if it were created by a human. This means that using an AI tool to sharpen an image doesn’t require disclosure, but using an AI tool to generate one part of an image does. An author will then receive copyright protection in only their contributions to the work and the changes they made to the AI-generated portions. As Thaler shows, an author must make some changes to an AI-generated work in order to receive any copyright protection at all in that work.

All of this means, broadly speaking, that the more an author changes an AI-generated work—such as by using tools like photoshop to alter an image or by editing AI-generated text—the more likely it is that the work will be copyrightable, and, by the same token, the less “thin” any copyright protection in the work will be. While there are open questions about how much creativity is required from a human in order to transform an AI-generated work into a copyrightable work of authorship, this case has underscored that at least some creativity is required—and using an AI tool that you yourself developed to create the work does not cut it. 

The way Thaler framed his Creativity Machine as the creator of the work in question also shows that it is important to avoid anthropomorphizing AI systems—just as the court rejected the notion of an AI-generated work being a work-made-for-hire, a creative work with both generative AI and human contributions probably could not be registered as a “co-authored” work. Humans are predisposed to attribute human characteristics to non-humans, like our pets or even our cars, a phenomenon which we have seen repeatedly in the context of chat bots. Regardless, it’s important to remember that a generative AI program is a tool based on a model. And thinking of generative AI programs as creators rather than tools can distract us from the established and undisturbed principle in copyright law that only a human can be considered an author, and only a human can hold a copyright. 

UPDATE: On Monday, August 14th, Judge Koeltl issued an order on the proposed judgement, which you can read here, and which this blog post has been updated to reflect. In his order, the judge adopted the definition of “Covered Book” suggested by the Internet Archive, limiting the permanent injunction subject to an appeal to only those books published by the four publisher plaintiffs that are available in ebook form.

After months of deadline extensions, there is finally news in Hachette Books v. Internet Archive, the case about whether Controlled Digital Lending is a fair use, which we have been covering since its inception over two years ago, and in which Authors Alliance filed an amicus brief in support of Internet Archive and CDL. On Friday, August 11th, attorneys for the Internet Archive and a group of publishers filed documents in federal court proposing “an appropriate procedure to determine the judgment to be entered in this case,” as Judge John G. Koeltl of the Southern District of New York requested. 

In a letter to the court, both parties indicated that they had agreed to a permanent injunction, subject to an appeal by IA, “enjoining the Internet Archive [] from distributing the ‘Covered Books’ in, from or to the United States electronically.” This means that the Internet Archive has agreed to stop distributing within the U.S. the books in its CDL collection which are published by the plaintiff publishers in the case (Hachette Book Group, HarperCollins, Wiley, and Penguin Random House), and are currently available as ebooks from those publishers. The publishers must also send IA a catalog “identifying such commercially available titles (including any updates thereto in the Plaintiffs’ discretion), or other similar form of notification,” and “once 14 days have elapsed since the receipt of such notice[,]” IA will cease distributing CDL versions of these works under the proposed judgment.

Open Questions

Last week’s proposed judgment did leave an open question, which Judge Koeltl was asked to decide before issuing a final judgment: should IA be enjoined from distributing CDL versions of books published by the four publishers where those books are available in any form, or should it only be enjoined from distributing CDL versions of these books that are available as ebooks? This difference may seem subtle, but it’s actually really meaningful. 

The publishers asked for a broader definition, whereby any of their published works that remain in print in any form are off the table when it comes to CDL. The publishers explain in a separate letter to the court that they believe that it would be consistent with the judgment to ban the IA from loaning out CDL versions of any of the commercially available books they publish, whatever the format. They argue that it should be up to the publishers whether or not to issue an ebook edition of the work, and that even when they decide not to do so (based on an author’s wishes or other considerations), IA’s digitization and distribution of CDL scans is still infringement. 

On the other hand, the Internet Archive is asked the judge to confine the injunction to books published by the four publishers that are available as ebooks, leaving it free to distribute CDL scans of the publishers’ books that are in print, but only available as print and/or audio versions. It argues that to forbid it from lending out CDL versions of books with no ebook edition available would go beyond the matters at issue in the case—the judge did not decide whether it would be a fair use to loan out CDL versions of books only available in print, because none of the works that the publishers based the suit upon were available only as print editions. Furthermore, IA explains that other courts have found that the lack of availability of a competing substitute (in this case, an ebook edition) weighs in favor of fair use under the fourth factor, which considers market competition and market harm.

It seems to me that the latter position is much more sensible. In addition to CDL scans of books only available as physical books not being at issue in the case, the fair use argument for this type of lending is quite different. One of the main thrusts of the judge’s decision in the case was his argument that CDL scans compete with ebooks, since they are similar products, but this logic does not extend to competition between CDL scans and print books. This is because the markets for digital versions of books and analogue versions of books are quite different. Some readers strongly prefer print versions of books, and some turn to electronic editions for reasons of disability, physical distance from libraries or bookstores, or simple preference. While we believe that IA’s CDL program is a fair use, its case is even stronger when it comes to CDL loans of books that are not available electronically. 

Then on Monday, August 14th, Judge Koeltl issued an order and final judgment in the case, agreeing with the Internet Archive and limiting the injunction to books published by the four publishers which are available in ebook form. Again, this may seem minor, but I actually see it as a substantial win, at least for now. While even the more limited injunction is a serious blow to IA’s controlled digital lending program, it does allow them to continue to fill a gap in available electronic editions of works. The judge’s primary reasoning was that books not available as ebooks was beyond the scope of what was at issue in the case, but he also mentioned that factor four analysis could have been different were there no ebook edition available.

Limitations of the Proposed Judgment

Importantly, the parties also stipulated that this injunction is subject to an appeal by the Internet Archive. This means that if the Internet Archive appeals the judgment (which it has indicated that it plans to do), and the appeals court overturns Judge Koeltl’s decision, for example by finding that its CDL program is a fair use, IA may be able to resume lending out those CDL versions of books published by the plaintiffs which are also available as ebooks. The agreement also does not mean that IA has to end its CDL program entirely—neither books published by other publishers nor books published by the publisher plaintiffs that are not available as ebooks are covered under the judge’s order.  

What’s Next?

The filing represents the first step towards the Internet Archive appealing the court’s judgment. As we’ve said before, Authors Alliance plans to write another amicus brief in support of the Internet Archive’s argument that Controlled Digital Lending is a fair use. Now that the judge has issued his final judgment, IA has 30 days to file a “notice of appeal” with the district court. Then, the case will receive a new docket in the Second Circuit Court of Appeals, and the various calendar and filing processes will begin anew under the rules of that court. We will of course keep our readers apprised of further developments in this case.