Generative AI lawsuits fall into three main categories: copyright claims over the material used to train AI models, right-of-publicity claims over cloned voices and likenesses, and defamation claims when AI systems fabricate false statements about real people. Courts have begun issuing substantive rulings, and they point in sharply different directions. One federal judge has rejected fair use for AI training that competed with the copyrighted source; another has embraced fair use when the training data was lawfully bought and the output served a different purpose. Billions of dollars in statutory damages are on the table, major record labels and news organizations are leading the charge, and Congress is weighing federal legislation on digital replicas. None of the biggest cases has reached an appellate court yet.
Copyright Claims Over Training Data
The largest cluster of suits targets the foundational step in building any large language model or image generator. To learn how language and images work, AI systems ingest enormous datasets scraped from the open internet, including novels, news articles, photographs, and illustrations. Copyright owners argue that this scraping creates unauthorized digital copies of their work, violating their exclusive rights to reproduce and distribute it under federal law.1Office of the Law Revision Counsel. 17 U.S. Code 106 – Exclusive Rights in Copyrighted Works Without permission or payment, these creators say, AI developers are building billion-dollar products on the back of stolen labor.
The financial exposure is staggering. Statutory damages for willful copyright infringement range from $750 to $150,000 per work.2Office of the Law Revision Counsel. 17 USC 504 – Remedies for Infringement: Damages and Profits When a single training dataset can contain millions of copyrighted images or documents, the potential liability quickly reaches the billions.
Universal Music, Concord Music Group, and ABKCO filed a $3 billion lawsuit against Anthropic in 2025 covering more than 20,000 copyrighted musical works, alleging that the Claude model was trained on pirated copies of lyrics and compositions. Sony Music, Universal Music, and Warner Records separately sued AI music startups Suno and Udio in 2024, claiming these platforms were trained directly on copyrighted recordings. Suno has since settled with Warner, and Udio has signed licensing agreements with Warner, Universal, and the independent label Merlin. Litigation between Sony and Udio continues.
The New York Times filed one of the highest-profile suits, arguing that OpenAI scraped decades of paywalled journalism to train ChatGPT and that the model can sometimes reproduce near-verbatim passages of Times reporting. A federal judge denied OpenAI’s motion to dismiss the core copyright claims. Discovery has shown that ChatGPT sometimes memorizes and regurgitates training data rather than merely “learning” from it, a technical reality that could undermine the argument that AI training is fundamentally different from copying. In Authors Guild v. OpenAI, the court ordered the company to turn over internal communications about why it deleted two major training datasets.3Justia Law. Authors Guild et al v. OpenAI Inc. et al
Visual artists have their own class action. In Andersen v. Stability AI, a group of illustrators sued Stability AI and Midjourney, alleging that the companies copied billions of images to train their image generators. The court allowed the direct copyright infringement claim to proceed, finding that the act of copying images into a training dataset, regardless of what comes out the other end, can itself constitute infringement. Additional claims that the models’ outputs are infringing derivative works also survived dismissal.
The Fair Use Split
AI companies almost universally argue that training on copyrighted material qualifies as fair use. Federal law requires courts to weigh four factors: the purpose and character of the use, the nature of the copyrighted work, the amount used relative to the whole, and the effect on the market for the original.4Office of the Law Revision Counsel. 17 USC 107 – Limitations on Exclusive Rights: Fair Use Early rulings have reached opposite conclusions.
The first major ruling went against AI. In Thomson Reuters v. Ross Intelligence, a Delaware federal court granted summary judgment to Thomson Reuters, holding that Ross’s use of Westlaw headnotes to train a competing legal-research AI was not fair use.5U.S. District Court for the District of Delaware. Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc. The court found the use was not transformative because Ross built a product that directly competed with the original, and it emphasized that even a potential market for AI training data weighs against fair use. The judge distinguished prior intermediate-copying cases involving software code, reasoning that headnotes are expressive works whose ideas can be reached without copying their specific language.
Then, in June 2025, a Northern California court reached the opposite conclusion. In a case involving Anthropic, Judge William Alsup ruled that copying lawfully purchased books to train an AI model is fair use, calling the transformation “spectacular.” The court reasoned that the AI learned from the works the way an aspiring writer learns from reading, not to replicate them but to create something fundamentally different. The same judge ruled that training on pirated copies of those same works was not fair use, because obtaining the material through infringement undermined the transformative purpose. The split within a single opinion signals that how a company acquires its training data may matter as much as what it does with it.
Fair use for AI training remains genuinely unsettled. The Thomson Reuters case suggests that when the AI product competes in the same market as the copyrighted material, fair use is a hard sell. The Anthropic ruling suggests that when training data is lawfully obtained and the AI’s output serves a different purpose, the defense can hold. Neither case has reached an appellate court yet, and the ultimate resolution could come from the Supreme Court.
When the Output Looks Like the Original
Even if courts eventually bless AI training as fair use, a separate question remains: what happens when the AI’s output closely resembles existing copyrighted work? Plaintiffs argue that the output itself is an infringing reproduction or derivative work. Courts evaluate this through substantial similarity, asking whether specific expressive elements were copied and whether an ordinary audience would recognize the resemblance.
Generative models do sometimes reproduce recognizable chunks of their training data. The New York Times case includes exhibits of ChatGPT outputting near-verbatim excerpts from published articles. Image generators have produced outputs containing faint traces of watermarks from stock photo services, suggesting the model retained specific pixel-level features from training images rather than abstracting general concepts.
AI developers counter with several defenses. The de minimis doctrine holds that copying so small it doesn’t implicate the copyright holder’s interests isn’t actionable. Another defense is that the “style” of an artist isn’t copyrightable: copyright law protects specific expressions, not broader techniques, color palettes, or literary voice. A prompt asking for an image “in the style of” a living painter doesn’t necessarily produce an infringing copy, even if the result looks like something that painter might have made. The distinction between protectable expression and unprotectable style is exactly where the hardest cases live, and the ability of AI to mimic a specific professional’s aesthetic cheaply and at scale raises the economic stakes.
Metadata Stripping Claims
A recurring technical claim targets the automated stripping of metadata during scraping. Federal law prohibits intentionally removing or altering copyright management information—data like the author’s name, the title of the work, and the terms of use—when the removal facilitates infringement.6Office of the Law Revision Counsel. 17 U.S. Code 1202 – Integrity of Copyright Management Information Plaintiffs allege that AI companies’ scraping tools automatically strip watermarks, bylines, and embedded metadata from images and text before feeding them into training pipelines, making it nearly impossible for creators to prove their work was used.
Statutory damages for these violations are separate from standard copyright infringement and range from $2,500 to $25,000 per violation.7Office of the Law Revision Counsel. 17 USC 1203 – Civil Remedies The music publishers’ lawsuit against Anthropic includes a dedicated claim under this provision, alleging that the company systematically removed identifying information from copyrighted lyrics and compositions. The challenge for plaintiffs is proving intent: they must show the developer knew the information was being stripped and understood it would facilitate infringement, not merely that the scraping tool happened to discard metadata as a technical byproduct.
Scraping, Terms of Service, and Access Controls
Alongside copyright claims, some plaintiffs have sued AI companies for breaching website terms of service that prohibit automated scraping. These contract-based claims have run into a significant obstacle: copyright preemption. In X Corp. v. Bright Data, a Northern California court dismissed breach-of-contract claims over scraping, holding that allowing platforms to enforce private no-scraping rules through contract law would create a shadow copyright system that conflicts with the one Congress enacted.
That setback has pushed plaintiffs toward a different legal theory. Rather than arguing that AI companies violated the fine print, they now allege that scraping tools bypassed actual technical barriers—CAPTCHAs, login walls, IP-rate limits, and other access controls—in violation of the DMCA’s anti-circumvention provisions. Reddit has filed lawsuits using this approach, and YouTube content creators have brought similar claims against Nvidia for allegedly evading access restrictions to scrape video content. For publicly available data with no authentication barrier, the prevailing view is that scraping alone does not violate the Computer Fraud and Abuse Act, a position anchored by the Ninth Circuit’s ruling in hiQ Labs v. LinkedIn.
Voice and Likeness Lawsuits
Generative AI can now clone a specific person’s voice or face with startling accuracy. This has opened a second front of litigation under the right of publicity, a body of state law that gives individuals control over commercial use of their name, image, voice, and likeness. Unlike copyright, which protects works, the right of publicity protects identity itself, and it varies significantly from state to state in scope and duration.
The most high-profile dispute involved Scarlett Johansson and OpenAI. After Johansson declined an offer to voice the ChatGPT assistant, OpenAI released a voice called “Sky” that many listeners found strikingly similar to her. Johansson publicly objected, and OpenAI pulled the voice. The incident echoed a 1988 Ninth Circuit ruling in Midler v. Ford, where the court held that using a sound-alike of Bette Midler’s voice in a commercial, after Midler had turned down the deal, was an actionable misappropriation under California law. Voice actors have also filed a class action against AI voice company Lovo, alleging that it trained on voiceover recordings available on freelance platforms and created unauthorized replications of performers’ voices.
SAG-AFTRA has pushed back on the labor side. In May 2025, the union filed an unfair labor practice charge against Llama Productions for allegedly using AI to recreate the voice of Darth Vader in a video game without bargaining with the voice actors who previously performed that work. The charge was withdrawn after a contract settlement, but it signaled that unions view AI voice cloning as a collective bargaining issue, not just an individual legal claim.
Congress has introduced the NO FAKES Act, which would create a federal intellectual property right over every individual’s voice and likeness, including after death, and allow lawsuits against anyone who knowingly creates, distributes, or profits from unauthorized digital replicas.8Congress.gov. S.1367 – NO FAKES Act of 2025 As of mid-2026, the bill has been referred to the Senate Judiciary Committee but has not yet received a floor vote in either chamber. If passed, it would fill a gap that currently leaves many individuals without recourse, since right-of-publicity protections vary widely and some states offer almost none.
Defamation From AI Hallucinations
A separate category of litigation has nothing to do with training data and everything to do with what AI models say once they’re deployed. Generative systems sometimes hallucinate, producing confident, specific, and completely fabricated claims about real people. When those fabrications are defamatory, the targets are starting to sue.
Defamation law requires a plaintiff to prove a false statement of fact, publication to a third party, fault on the defendant’s part, and resulting harm to reputation. The fault standard varies depending on who the plaintiff is. Public officials and public figures must prove actual malice, meaning the defendant knew the statement was false or acted with reckless disregard for the truth, a standard set by the Supreme Court in New York Times v. Sullivan. Private individuals generally need to show only negligence.
Both standards create problems when the speaker is an algorithm. In Walters v. OpenAI, the court granted summary judgment to OpenAI after finding no evidence of negligence or actual malice, in part because the company had provided warnings about potential inaccuracies in its outputs. In Battle v. Microsoft, the plaintiff’s case hinges on showing that Microsoft failed to exercise reasonable care over the information its AI distributed. These early cases suggest that AI companies can reduce their defamation exposure by disclosing the limitations of their systems, though that defense has obvious limits as AI tools become embedded in products that users treat as authoritative, like search engines and workplace assistants.
The application of Section 230 of the Communications Decency Act adds another layer of uncertainty. That law protects internet platforms from liability for content provided by another information content provider, meaning user-generated content.9Office of the Law Revision Counsel. 47 U.S. Code 230 – Protection for Private Blocking and Screening of Offensive Material AI-generated text is not user-generated; the model produces it. A Congressional Research Service analysis has noted that courts have not yet decided whether Section 230 applies to generative AI outputs at all.10Congress.gov. Section 230 Immunity and Generative Artificial Intelligence If courts determine that AI companies are the speakers of their models’ output rather than passive intermediaries, Section 230 would offer no shield, and defamation liability could expand dramatically.
Who Owns What the AI Makes
A related question concerns what comes out of these systems and who, if anyone, owns it. The U.S. Copyright Office has consistently held that copyright protection requires human authorship.11U.S. Copyright Office. Copyright and Artificial Intelligence In its 2023 registration guidance and subsequent decisions, the Office has refused to register works created entirely by AI, including a visual artwork generated by the “Creativity Machine” system and AI-generated images in the comic Zarya of the Dawn. Hand-selected arrangements of AI images can qualify. A review board explicitly rejected the argument that AI output could be treated as a “work made for hire,” reasoning that an AI system cannot be an employee or enter into a contract.
For anyone using AI tools professionally, the practical consequence is significant: purely AI-generated output may have no copyright protection at all, meaning anyone could copy it freely. Works that involve meaningful human creative choices—selecting, arranging, and modifying AI outputs—can qualify for registration, but the line between enough human involvement and not enough remains blurry. The Copyright Office issued Part 2 of its report on copyright and AI in January 2025, and the framework is still actively evolving.
Where Things Stand
The generative AI litigation landscape is moving fast but remains far from settled. The two most significant fair use rulings point in opposite directions, one rejecting the defense when an AI product directly competes with the copyrighted source, the other embracing it when lawfully obtained training data produces something fundamentally different. Neither has been reviewed on appeal. The major copyright cases against OpenAI, Stability AI, and Midjourney are deep in discovery but years from trial. The music industry has shown that licensing deals can emerge from litigation pressure, with multiple AI startups signing agreements with labels rather than risking a jury verdict. Congress is considering federal protections for digital replicas but hasn’t passed anything yet. Whether AI companies bear liability for their models’ false statements about real people remains essentially untested. The answers that emerge over the next few years will set the rules for an entire industry.