[ACC Note: The parties in Bartz v. Anthropic notified the court on August 26, 2025, that they have reached a settlement. Court approval of the settlement is pending.]
Fair use is gaining traction as a legal basis for artificial intelligence vendors to train AI models using third-party content - at least in the United States.
In June 2025, two federal judges in California ruled partially in favor of AI companies Anthropic and Meta in lawsuits brought by groups of authors. The judges found that using copyrighted books to train generative AI models like Claude and Llama 2 can qualify as a permissible “fair use” under US copyright law. However, these decisions do not give AI developers a free pass regarding the methods they use to acquire content.
In-house counsel should be aware of potential implications for their businesses. Below are key takeaways. “Taken together, the Anthropic and Meta decisions reflect the evolving landscape of copyright law in the GenAI era,” said Joe Petersen, a partner with Kilpatrick Townsend & Stockton LLP. “While both decisions affirmed the fair-use defense on their respective records, they underscore that success in future cases—whether by rightsholders or technology companies—will hinge on the specific facts, the development of the evidentiary record, and the courts’ willingness to engage with the unique market implications posed by generative AI,” shared Petersen.
Also check out the new ACC AI Center of Excellence for In-house Counsel, a powerful resource to help in-house lawyers build AI capabilities in four core areas:
- Your role as in-house lawyer: Build the knowledge and confidence to understand and apply AI in your in-house legal work.
- Your law department: Discover practical ways your legal team can evaluate, adopt, and get the most out of AI tools.
- Your support of the enterprise: Play a key role in shaping your company’s AI strategy—helping balance innovation, risk, and governance.
- Your legal profession: Stay ahead of the curve on the ethical, professional, and regulatory questions AI brings to the legal world.
AI Training May Be Fair Use
The judges found that the AI training in these cases was a fair use, but they offered different reasons for their findings.
Both decisions go over four key factors that play into determining whether a use is fair under the US Copyright Act:
Factor 1: the purpose and character of the use, including whether such use is of commercial nature or is for nonprofit educational purposes;
Factor 2: the nature of the copyrighted work;
Factor 3: the amount and substantiality of the portion used in relation to the copyrighted work as a whole; and
Factor 4: the effect of the use upon the potential market for or value of the copyrighted work.
- Bartz v. Anthropic. The judge highlighted that the AI training process was highly "transformative," which strongly supported a finding of fair use. This is because the purpose of the training was to teach the Claude model to generate new, unrelated text - not to reproduce the content of the books on which it was trained.
- Kadrey v. Meta. While agreeing that AI training is transformative, the judge in Kadrey based his decision in favor of the fair use argument on the authors' failure to provide evidence that using their books for training would harm the market for those books. The judge also noted a more expansive theory of “market dilution”—where a flood of AI-generated content could devalue original creative works like those by the plaintiff authors—but did not base his decision on it, as the plaintiffs had not provided evidence to support it. The decision strongly hints that if the plaintiffs had developed a record of evidence showing such market dilution, the judge may have rejected the defendant’s fair use defense.
Beware of Piracy
These decisions do not mean that AI developers can use content that they acquire in ways that infringe others’ rights.
In the Bartz v. Anthropic case, while accepting the fair use argument for legally acquired training data, the court also ruled that the company could not claim fair use for books it allegedly downloaded illegally from pirate sites to help create its permanent library of training data. Anthropic allegedly decided ultimately not to use (or continue using) the pirated books to train Claude, but it retained its copies of those books in its library.
This aspect of the Anthropic case is proceeding to trial and could result in significant damages against Anthropic, which allegedly downloaded over seven million books from pirate sites. Keep in mind this ruling was on a motion for partial summary judgment only on the fair use argument. Monitor the developments of this case, as its outcome will depend on the court’s evaluation of a likely fuller set of facts.
Five Key Takeaways for In-house Lawyers
- Proceed with Caution on AI Training. The legal rules for using copyrighted materials in AI training are still taking shape, and these court decisions are not the final word. Other federal courts may still weigh in, and appeals are likely. Also, many countries do not recognize a "fair use" defense for copyright violations, meaning conduct that is permissible in the US may be illegal elsewhere.
- Protect Your Organization’s Intellectual Property. Work with others across the business, such as IT, InfoSec, and Web developers, to ensure adequate safeguards against pirating and other types of unauthorized downloading, such as web scrapers. Proactive protection of your organization’s content is a precaution that can save you the headache of later on having to bring claims of copyright infringement due to unwanted content scraping.
- Vet Your Vendors. When buying AI-enabled tools or services, ask how the models were trained. If a model’s training violated copyright laws, it may later require retraining, might suffer from downgraded performance, or may become unavailable altogether due to restrictions (for example as the result of litigation by rights holders).
- Mind How Content Is Acquired. Even if AI training can be a fair use of copyrighted data, getting your own organization’s content through illegal means is not. These cases are a strong reminder to confirm that your business content, such as for creative products, marketing, or websites, is legally sourced. Ensure that marketing teams and other producers of content in your organizations are trained on basic rules, including the dos and don’ts of using or incorporating third-party content.
- Monitor AI Outputs. Businesses that use AI to generate public-facing content should be cautious not to publish AI outputs that violate others' copyrights. Where possible, consider using enterprise-level AI products that offer contractual protections, such as representations, warranties, and indemnification provisions. Be mindful of the restrictions and exclusions that the AI vendor may have built into such terms and conditions. And if you intend to generate content over which you want your organization to claim intellectual property rights, be mindful of legal requirements of human authorship that may apply under the relevant laws – using AI-generated output “as is” without any substantial transformation by a human author might not be sufficient to claim copyright protection over that content. Learn more with the checklist Protect Intellectual Property When Using Generative AI, in the ACC Artificial Intelligence Toolkit for In-house Lawyers, sponsored by Kilpatrick.
These cases show that courts are starting to shape how copyright law applies to generative AI. But many legal questions remain. In-house lawyers should stay up to date with legal and regulatory developments in this area as they work to protect their businesses and develop effective AI governance.
Keep up to date on AI legal and regulatory developments and get up to speed with AI for the in-house lawyer, the law department, the enterprise, and the legal profession with the ACC AI Center of Excellence for In-house Counsel.