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.