If you've been paying attention to AI litigation over the past year, you already know the basic shape of the argument: AI companies scraped the internet, used millions of copyrighted works without permission or payment, built products worth billions of dollars on that foundation, and are now defending that practice as legally protected. What's changed recently is the intensity. Artists, authors, and major publishers have stopped treating these lawsuits as long-shot nuisance claims and started treating them as existential fights worth serious legal firepower.
The cases piling up in federal courts right now aren't just about who gets a check. They're about whether the current model of building AI, which depends on ingesting vast quantities of human-created work, can survive contact with copyright law.
Who's Suing Whom, and Why Now
The list of plaintiffs has grown considerably. The New York Times filed its landmark case against OpenAI and Microsoft in December 2023, alleging that GPT models were trained on millions of its articles without a license. Authors including George R.R. Martin, John Grisham, and Jodi Picoult joined a class action against OpenAI through the Authors Guild. Visual artists have sued Stability AI, Midjourney, and DeviantArt. Getty Images went after Stability AI in both U.S. and UK courts, producing some of the most striking evidence in any of these cases: AI-generated images that still contained mangled versions of the Getty watermark, which suggested the training data was unmistakably direct.
Universal Music Group, Sony Music, and Warner Music Group collectively sued Anthropic in October 2023 over lyrics appearing in Claude outputs. A separate case targets Suno and Udio, two AI music generation startups, with the Recording Industry Association of America seeking up to $150,000 per infringed work. That last number matters because at scale, the potential damages become almost impossible to calculate.
The timing isn't accidental. Several factors converged. AI companies started generating serious revenue, which made them worth suing. The outputs of these systems became good enough that the connection between training data and generated content became harder to dismiss. And critically, discovery in early cases began producing internal documents that plaintiffs found useful.
The Fair Use Defense Is Shakier Than It Looks
Every major AI company has anchored its legal defense in the doctrine of fair use, arguing that training a model on copyrighted work is transformative and therefore protected. Google successfully used a similar argument in the long-running Authors Guild v. Google case over book scanning, and companies like OpenAI are betting that logic carries over.
It might not. The Google Books case involved digitizing books to make them searchable, with only snippets shown in results. Generative AI does something fundamentally different: it produces new content that can directly substitute for the original works. A user who asks Claude to "write something in the style of a New York Times investigative piece" is arguably competing with the Times in a way that a search index does not.
Courts look at four factors in fair use analysis: the purpose and character of the use, the nature of the copyrighted work, how much of the work was used, and the effect on the market for the original. That fourth factor is where AI companies face the most exposure. Publishers and record labels can point to the fact that AI-generated content is already replacing commissions and licensing deals that would otherwise go to human creators. That's not a hypothetical harm.
Some legal scholars think the AI companies have a defensible position on the transformative use argument. Others point out that the companies used entire works, not excerpts, and that their commercial purpose is obvious. The honest answer is that existing copyright law wasn't written with any of this in mind, and no one knows how courts will ultimately rule.
What the Companies Are Actually Doing With Data
Training large language models requires enormous datasets. GPT-4 was reportedly trained on roughly 1 trillion tokens of text data. Models like Llama 2 from Meta used datasets including Common Crawl, which is essentially a snapshot of a large portion of the publicly accessible web. Common Crawl contains copyrighted news articles, books, academic papers, forum posts, and creative writing, all scraped without specific permission from individual rights holders.
Some companies have started signing licensing deals, partly as a legal hedge and partly because the supply of high-quality data is tighter than it used to be. The Associated Press signed a deal with OpenAI. Reddit licensed its content to Google for reportedly around $60 million annually. Axel Springer, which owns Politico and Business Insider, signed with OpenAI. Stack Overflow inked a deal with Google.
These deals signal that the industry knows the pure scrape-and-train approach has a shelf life. But they also reveal the gap between who gets a deal and who doesn't. A major media conglomerate has lawyers and leverage. An independent illustrator whose work was included in LAION-5B, the dataset used to train Stable Diffusion, has neither.
The Dataset Problem Is Bigger Than Anyone Wants to Admit
LAION-5B contained approximately 5.85 billion image-text pairs. Researchers have since identified personal photos, medical images taken without consent, and copyrighted artwork throughout that dataset. Spawning, a company that built the Have I Been Trained tool, found that artists could search for their work and opt out, but only after the training was already done. The concept of retroactive consent doesn't map cleanly onto how copyright is supposed to work.
Researchers at the University of Washington found that diffusion models can reproduce training images with high fidelity under certain prompting conditions, which directly undermines the argument that these models don't "store" copyrighted content in any meaningful sense. They do, in some functional way, and they can retrieve it.
For visual artists in particular, the situation has been radicalizing. Platforms like ArtStation saw massive protests in 2022 when artists discovered their work had been used without permission. The grievance isn't just financial, though that's real. It's that an industry worth hundreds of billions of dollars was built by treating an entire profession's life work as free raw material.
What a Loss Would Actually Mean for AI Development
If courts rule that training on copyrighted data without a license constitutes infringement, the consequences would cascade quickly. Companies would face potential statutory damages across millions of works, damages that could be ruinous even for well-capitalized firms. They'd need to either purge infringing data from existing models (which may not even be technically possible in a clean way) or negotiate licenses with rights holders at scale.
That second option would restructure the industry significantly. Licensing deals would favor large publishers and studios with the legal capacity to negotiate, further marginalizing individual creators. Startups without the resources to pre-clear training data would struggle to compete with incumbents that already have models trained on data they're now retroactively licensing. The barrier to entry for building a frontier model would get much higher.
Synthetic data generation has emerged as one possible path forward. Companies are already training models on outputs from other AI systems, which sidesteps some of the copyright issues. But the quality of models trained primarily on synthetic data is still an open question, and there's a real risk of compounding errors when systems learn from each other's outputs rather than from primary human-created sources.
What This Means for the People Actually Creating Things
For working writers, artists, and musicians, the litigation wave represents the first serious attempt to establish that their work has value in the AI supply chain. Even if individual plaintiffs don't win large verdicts, the cases are already pressuring companies to negotiate rather than simply take. That's a meaningful shift from two years ago.
A licensing regime, if it develops, would probably flow money through collective management organizations rather than directly to individual creators, similar to how music royalties work through ASCAP or BMI. That model has well-known problems with transparency and fair distribution, but it beats the current situation where most creators receive nothing.
The broader cultural stakes are harder to quantify. If generative AI makes it economically unviable to be a professional illustrator, novelist, or session musician, the pipeline of human-created work that these models depend on starts to thin. There's a version of this story where the industry's refusal to compensate creators ends up undermining the quality of the very systems they're building. Whether that feedback loop plays out over years or decades is unclear, but it's not a theoretical concern.
The courts haven't resolved any of this yet, and Congress hasn't moved meaningfully on AI and copyright despite multiple hearings. What's already resolved is that the era of building AI on uncontested free data is ending, one lawsuit at a time.
