April 15, 2025 AI Training Data: The Copyright Controversy

As generative AI progresses, so do lawsuits over its use of copyrighted data. Here’s where key legal battles on infringement and fair use stand.

By Laurel Gilbert Rogowski & Liz Nagle

Copyright Infringement Theories Asserted as to Generative AI

“Generative AI” is a type of AI that produces new content including text, images, audio, and/or video.  Many of the leading copyright infringement cases regarding generative AI relate to “large language models” (“LLM”), which are trained on vast datasets and utilize deep learning to generate human-language text outputs.  In cases currently pending throughout the country, plaintiffs have raised several theories as to how LLMs may infringe copyrights in the materials used to train them, to varying degrees of success. 

Direct Copyright Infringement

In numerous pending actions, owners of copyrighted works have alleged that the unauthorized use of their copyrighted materials to train LLMs constitutes direct copyright infringement.  Many such plaintiffs also allege that the LLMs’ output of text that is identical to or derivative of the copyrighted works used to train the model is an additional form of direct copyright infringement.

These claims have largely survived the early dismissal stage, and defendants have typically challenged their viability only on narrow procedural grounds—for example, arguing that the claims are untimely.  These claims will likely ultimately rise or fall on the defendants’ fair use defenses,¹ discussed further below.

Contributory Copyright Infringement

A defendant may also be liable for contributing to another’s direct copyright infringement if the defendant knew of the infringing activity and induced, caused, or materially contributed to it.  Many copyright owners have alleged that LLM owners materially contributed to copyright infringement by end users of the generative AI products, with knowledge that the LLMs were capable of distributing unauthorized copies or derivatives of the plaintiffs’ copyrighted works.  These claims have also largely survived dismissal.  For instance, in Andersen, the court rejected dismissal of such a claim where the plaintiffs’ allegations plausibly supported an inference that the generative AI platform “by operation by end users creates copyright infringement and was created to facilitate that infringement by design.”²  In late March, in the leading case New York Times Co. v. Microsoft Corp., the court similarly denied the defendants’ motion to dismiss a claim alleging that they contributed to copyright infringement by end-users of OpenAI’s ChatGPT products.³

The Digital Millennium Copyright Act (“DMCA”)

A third form of copyright claim common in generative AI litigation is claims for alleged violations of the DMCA.  The DMCA makes it unlawful intentionally to remove, alter, or falsify copyright management information (“CMI”), including a copyrighted work’s title, author, and copyright notice, or to distribute a work knowing that CMI has been removed or altered, where the defendant knows or has reasonable grounds to know that such removal, alteration, or distribution will induce, enable, facilitate, or conceal copyright infringement.

Claims alleging DMCA violations based on the removal of CMI when building an LLM training dataset containing copyrighted works have met varying success to date.  In a November 2024 decision, a court dismissed claims that OpenAI had violated the DMCA by removing CMI from the plaintiffs’ works used to train ChatGPT based on the plaintiffs’ failure to establish standing by alleging a concrete, particularized injury resulting from the alleged CMI removal.⁴ In February 2025, a different judge of the same court rejected a similar standing argument in The Intercept Media, Inc. v. OpenAI, Inc.,⁵ finding that the plaintiff had sufficiently alleged harm to its property rights.  The Intercept Media court also concluded that the plaintiffs had pled cognizable DMCA violations by plausibly alleging that OpenAI intentionally removed CMI from the plaintiffs’ copyrighted works with knowledge that doing so would facilitate copyright infringement by end-users of ChatGPT.  The court in the New York Times matter also recently dismissed some DMCA claims while allowing others to proceed,⁶ underscoring that the viability of such claims depends upon the specific facts alleged.

Fair Use and the Reuters Decision

Perhaps the most closely watched issue in generative AI litigation is whether the unauthorized use of copyrighted works in the development of LLMs constitutes fair use.  Fair use is an affirmative defense to copyright infringement claims.  The four factors that guide a court’s analysis of whether an unauthorized use of a copyrighted work is “fair” are: (1) purpose and character of the use, (2) nature of the copyrighted work, (3) substantiality of the portion used, and (4) effect of the use on the copyrighted work’s value or potential market.

While defendants in generative AI cases are likely to put forth arguments on all four fair use factors tailored to the specifics of each case, there will be a particular focus on factors 1 and 4.  A recent, much-discussed decision from a Delaware federal court in Thomson Reuters v. ROSS Intelligence,⁷ after stating that these factors “weigh most heavily in the analysis,” granted summary judgment that using copyrighted “headnotes” to train a competitor’s AI system was not fair use.  Under factor 1, the court found that the defendant’s use of the copyrighted works to develop a competing legal research tool was not “transformative” because it did not have a “further purpose or different character” from the plaintiff’s.  Factor 4 also favored the plaintiff because the defendant’s use amounted to market substitution.  Notably, the Reuters decision carefully pointed out that the AI system at issue was “not generative AI”– and defendants in the generative AI cases will most certainly distinguish it on this basis.

Indeed, in recent weeks, the court in a case brought against Anthropic, creator of the “Claude” LLMs, distinguished the Reuters decision from issues before it at the preliminary injunction stage on the grounds, among others, that Reuters did not concern a generative AI model and the parties in that case were direct competitors.⁸  In a separate case, Meta distinguished Reuters in support of its argument that its use of copyrighted books to train its “Llama” LLMs was transformative and thus warrants summary judgment of fair use.⁹

Litigants are beginning to test the boundaries of well-established intellectual property laws against this incredibly dynamic technology.  Expect decisions this year that will have major consequences—not only for how AI tools are built and trained, but also for the rights of creators in the digital age. As the courts continue to wrestle with these complex issues, one thing is clear: the intersection of AI and copyright law will remain a battleground for innovation and the future of creative ownership.

About the Authors:

Laurel Gilbert Rogowski is a Partner in Hinckley Allen’s Litigation Group. Laurel’s practice is focused on intellectual property litigation. Her experience includes representing clients in disputes involving trademarks, copyright, patents, trade secrets, and unfair competition, as well as a wide range of commercial disputes, including partnership and contract disputes, business torts, cybersecurity and data privacy disputes, and product liability and premises litigation.

Liz Nagle is an Associate in Hinckley Allen’s Litigation group, with a focus on intellectual property disputes

¹ For example, in Andersen v. Stability AI Ltd., 3:23-cv-00201-WHO (N.D. Cal.), Dkt. 223, the court noted in denying a motion to dismiss the plaintiffs’ direct copyright infringement claim that whether the use of the copyrighted works in either operation or output was “substantial enough” to qualify for a fair use defense would need to be tested on summary judgment.

² Andersen, 3:23-cv-00201-WHO, Dkt. No. 223, at 9.

³ 23-cv-11195 (SHS) (OTW) (S.D.N.Y.), Dkt. No. 485; see also Dkt. No. 514. 

⁴ Raw Story Media, Inc. v. OpenAI, Inc., 24 Civ. 01514 (S.D.N.Y.), Dkt. No. 117.

⁵ 24-cv-1515 (JSR) (S.D.N.Y.), Dkt. No. 127.

⁶ New York Times Co., 23-cv-11195 (SHS) (OTW), Dkt. No. 485.

⁷ 1:20-cv-00613 (D.Del.), Dkt. No. 772.  On April 4, the court granted Defendant’s motion to certify this decision for interlocutory appeal to the Third Circuit.  Dkt. No. 799.

⁸ Concord Music Group v. Anthropic, 5:24-cv-03811-EKL (N.D. Cal.), Dkt. No. 321.

⁹ Kadrey, et al. v. Meta Platforms, Inc., 3:23-cv-03417-VC (N.D. Cal.), Dkt. No. 489.

Subscribe to Industry Today