Can Gen AI and Copyright Coexist?

According to a recent industry report, the “core” creative industries—notably, those that produce books, entertainment software, periodicals, movies, recorded music, television, and video games—contributed $1.8 trillion to the U.S. GDP in 2021. That’s nearly 8% of the U.S. economy. So it’s a big deal, as has been widely reported, that gen AI companies have been using unlicensed copyrighted materials from those industries to train their models. Already industry leaders and academic research say that gen AI tools have replaced workers in other important areas of the economy, and it’s not hard to imagine that it might do the same for creative artists. Which is why rightsholders in these industries have filed more than 40 cases in federal courts against gen AI companies for training their models on unlicensed copyrighted materials. Two weeks ago, two courts in the Northern District of California handed down the first two decisions on these cases—and neither decision was particularly positive for rightsholders. In the first case, Bartz v. Anthropic, which involved a suit brought against Anthropic by three authors, Judge William Alsup determined that training gen AI models with copyrighted material is fair use, and thus can be done without compensating rightsholders. “Like any reader aspiring to be a writer,” Judge Alsup wrote, “Anthropic’s LLM [is] trained upon works not to race ahead and replicate or supplant them—but to turn a hard corner and create something different.” In the second case, Kadrey v. Meta, which involved a suit brought against Meta by 13 authors, Judge Vincent Chhabria took a different position, finding that the unlicensed use of copyrighted materials for training is not likely to be a fair use, because LLM-generated content is fundamentally different from human-generated creation. “Using books to teach children to write,” Judge Chhabria reasoned, “is not remotely like using books to create a product that a single individual could employ to generate countless competing works with a miniscule fraction of the time and creative it would otherwise take.” We’re experts in the study of digital transformation and have given this issue a lot of thought. We recently served, for example, on a roundtable of 10 economists convened by the U.S. Copyright Office to study the implications of gen AI on copyright policy. We recognize that the two decisions are far from the last word on this topic; both will no doubt be appealed to the Ninth Circuit and then subsequently to the Supreme Court. But in the meantime, we believe there are already many lessons to be learned from these decisions about the implications of gen AI for business—lessons that will be useful for leaders in both the creative industries and gen AI companies. Lessons for Rightsholders As a rightsholder in the present environment, you can take several steps to protect your business. Recognize it’s possible to remove your copyrighted material from the training corpus. It’s tempting to think that the battle has already been lost: Once a piece of copyrighted content enters an LLM’s training corpus, there’s no way to get it out, right? Actually, wrong. When LLMs release major new versions (as in, say, the move from ChatGPT 3 to ChatGPT 4), they typically retrain the new model on the full training corpus. In that moment, there’s an opportunity for the LLM to remove a rightsholder’s copyrighted material from their training corpus—or for the rightsholder to obtain a court order instructing them to do so. Enforce rights against gen AI companies that are using your copyrighted materials without a license. If you believe your content is being used to train LLMs with a proper license, now might also be a good time to join the list of 40+ plaintiffs who are suing in court to protect their legal rights. This is particularly true given that some of the recent court decisions provide what amounts to an outline of how future plaintiffs could successfully argue their cases: Convince the court that gen AI models could be used to destroy the livelihoods of the very artists who are being asked to train those models for free. That’s a high bar, but at least it’s something. License where appropriate. Over 70 rightsholders—including HarperCollins, Universal Music, Reddit, Shutterstock, and the Wall Street Journal—have licensed their content for training by gen AI companies. If you haven’t yet licensed your content, now might be the time to consider pursuing licensing deals, possibly by creating curated datasets that are tailored for specific model training and thus would be attractive to AI developers. Your challenge in signing such licenses is that you are negotiating at a disadvantage given the direction the courts seem to be going with respect to fair use. But your advantage is timeliness and accuracy: Gen AI companies may be willing to pay to license content now rather than wait for these cases to make their way through the courts, and they may be willing to pay for a clean, reliable, and curated dataset rather than what they can get by indiscriminately scraping the web. Consider taking copyrighted content off the open web. Many IP businesses make some of their content available publicly in ways that drive their business models. There are cases across the spectrum of IP where content is made available “for free” to support advertising-based business models, as with the numerous ad-supported movie clips you can find on YouTube. Some companies also employ a “freemium-to-premium” model, in which consumers are offered some free content but need to pay for a more convenient channel or a more complete set of content. Think of the five to 10 free articles some websites give away every month before asking people to subscribe. Those strategies may have made sense before gen AI, but the rules have now changed, and it makes sense to rethink how you weigh the benefits of making your content available on the open web against the harms of your copyrighted material being used to train LLMs without your consent. This is particularly true given that some of the largest licensing deals have been for content that was protected behind a paywall, which made it difficult for gen AI companies to scrape. Protect copyrighted content on the open web. Speaking of web scraping, rightsholders should seek to protect any of their IP that remains on the open web. There are several avenues that can be pursued here, although none is perfect. Rightsholders can use a file called robots.txt to tell AI crawlers they shouldn’t scrape the site’s content. This will only work if LLMs choose to respect a site’s robots.txt file, but it’s a start. There are also ways for web hosting companies to block LLM scrapers, or to establish monetization policies for AI agents seeking to access subsets of your content. Lastly, there are emerging tools, such as Glaze or Nightshade, that can protect publicly available image-based IP from use in gen AI training. Lessons for Gen AI Companies If you’re a leader at a gen AI company, you might think you should just keep doing what you’re doing, given the way the courts have been ruling. But that might be a mistake. Here’s what you should be thinking about—and doing—as the courts sort things out. Factor in the financial risks. Although Judge Alsup found that the use of copyrighted data to train Anthropic’s LLM was fair use, he did so with a significant caveat: The copyrighted data could not have been obtained through piracy. Specifically, Alsup wrote that “piracy of otherwise available copies is inherently, irredeemably infringing even if the pirated copies are immediately used for the transformative use.” This means the use of pirated content carries significant financial risk. Under U.S. law (17 U.S.C. § 504), copyright holders can be granted statutory damages of up to $30,000 per work infringed, and up to $150,000 if the court finds the infringement was willful. Alsup found that Anthropic used 7 million pirated books in its training data and ordered the company to stand trial for “the pirated copies used to create Anthropic’s central library and the resulting damages, actual or statutory (including for willfulness).” Seven million books times $150,000 in potential damages per book is a lot of potential financial risk—$1.05 trillion to be exact. And this isn’t just an Anthropic issue. Discovery in Kadrey v. Meta found that Meta used at least 82 TB of pirated book data to train its models. Other lawsuits, among them Tremblay v. OpenAI and O’Nan v. Databricks, claim use of similar pirated “shadow libraries” for training. Sign licenses. Given these financial risks, now might be the right time to consider signing licenses with rightsholders—particularly since the recent court cases may make rightsholders more willing sellers in the short term. Generate opt-out tools, allowing rightsholders to filter or remove their content from training datasets. In the face of lawsuits from rightsholders, companies such as YouTube and Facebook developed tools giving those rightsholders the ability to remove their content from their platforms. Now might be a perfect time for gen AI companies to develop similar tools. Reconsider the value of gaining more unlicensed data versus the legal risks. Eleuther.ai, an AI-research organization, recently released an 8 TB dataset called Common Pile v0.1, which it claims is made up entirely of open source or licensed content. The company also claims that when it used this model to train its LLMs, those models performed as well as models using unlicensed copyrighted data. In a blog post, the company’s developers wrote that their research suggests that “the common idea that unlicensed text drives performance is unjustified.” We aren’t able to independently verify those claims, but given the potential legal risks generated by the use of unlicensed (and frequently pirated) copyrighted content, maybe it’s a good idea for gen AI companies to question whether the marginal benefits that come from using more unlicensed copyrighted material justify the additional legal and financial risks. . . . Gen AI has the potential to benefit industry and society in many ways. But achieving that potential will require more robust and transparent partnerships between technology firms and the creative industries. On our current path we risk killing the goose—or in this case the authors, musicians, coders, and filmmakers—who laid the golden eggs that are key to the present and future value of gen AI output.

Can Gen AI and Copyright Coexist?

Summary.

Partner Center