Is AI Built on Theft? Almost 100 Lawsuits Say So.

Developers of artificial intelligence have been hit with dozens of lawsuits brought by content creators who allege the companies stole their work to turn a profit and create products that may ultimately replace their industries.

The suits, brought by news outlets, authors, photographers, and songwriters, argue that the companies used their work to train and develop AI tools without compensating them, while at the same time racing to make their creative work obsolete.

Others are headed for summary judgment, where the judge will decide the matter without a full trial. 

More broadly, they underscore the tension between the benefits of developing new technology and the possible destruction of classic, human-driven information gathering and creative production.

The ‘Doom Loop’

The news companies and others argue that AI firms harm their enterprise by using their high-quality work—generated through time, effort, and money—without compensating the sources.

“Defendants seek to free-ride on The Times’s massive investment in its journalism by using it to build substitutive products without permission or payment,” a suit brought by The New York Times and more than a dozen news outlets said.

Newly revealed documents from that lawsuit—brought against OpenAI and Microsoft—seem to show that the companies were aware that their methods could be problematic.

“Millions of people around the world will soon consider large models ‘hoovering up’ all their work to be an astonishing theft of unprecedented proportions,” Brent Hecht, Microsoft’s Director of Applied Science, said, according to a legal filing. 

“Our AI content strategy has started a ‘doom loop’ that will hurt the performance of our models and the entire web at the same time,” Hecht said in another internal document cited in court filings.

Since 2023, OpenAI has worked to secure licenses from news outlets used in training its models, beginning with The Associated Press in July of that year. Deals with News Corporation, The Washington Post, Politico, and others followed.

The lawsuits come from companies that rejected those licensing agreements.

Meta made similar deals, but was likewise sued by individual authors with whom it made no agreements. Google partnered with social media giant Reddit to train its model, but only after a lawsuit by individual plaintiffs who said the company was scraping private users’ data from across the internet.

Anthropic has not announced any licensing deals, and has instead relied on the fair use doctrine to defend itself in court.

Fair Use?

In order to train their AI models to respond correctly to user prompts, companies must first feed those models with a vast amount of information. That information has to come from somewhere, and much of it is under copyright.

But the companies have argued that scooping up copyrighted text to train AI models is protected under the fair use doctrine of Section 107 of the Copyright Act.

Fair use allows exceptions to copyright depending on the circumstances. Some examples include using copyrighted material for research, teaching, or criticism.

Whether the exception applies depends on how the copyrighted material is being used, Kevin Frazier, director of AI innovation and law at the University of Texas–Austin, told The Epoch Times.

“The rationale behind that predominantly turns on the first factor that’s weighed by courts when they’re analyzing whether something falls under the fair use doctrine, and that’s this vague but incredibly important term of ‘transformational use,’” he said.

Cases in which AI companies “scrape” the internet, books, or other sources for text and use that data to train their models are more likely to fall under fair use, Frazier said.

“The resulting thing that gets produced is this highly complex, highly nuanced, highly sophisticated AI model that is wildly different from the original work that was used to train the model itself, and so for that reason, many people regard it as transformative.”

In 2025, Meta beat a lawsuit filed by 13 authors who said it used their work to train its AI model, “Llama.”

The judge in that case ruled that using their books to train the model was “highly transformative,” and that the plaintiffs did not show that Meta’s product had any impact on the market value of their work.

The U.S. Department of Justice weighed in on the fair use question in early September, arguing that AI is important to national security, and the government has a “sovereign interest” in promoting innovation in the field.

The government is asking courts to rule that the fair use doctrine lets AI companies harvest the information from copyrighted material. That’s necessary to stoke competition from smaller firms, the DOJ said, since “only the largest technology companies might have the capital necessary to pay licensing fees.”

Input Versus Output

Cases where the AI model outputs new material are less likely to pass muster, Frazier said.

“If you ask a model to develop a new chapter of ‘Harry Potter,’ or let’s just say a new chapter of a very famous book, well in this instance, it does not seem to be as transformative, right?” he said.

Courts will weigh other factors too, such as impact on potential market value of the copyrighted work, whether the material used was factual or creative, and what portion of the work the AI model used, according to Frazier.

According to one court filing, OpenAI’s head of ChatGPT wrote that AI products pose an “existential threat” to publishers, “are largely substitutive, period,” and “will get more and more substitutive as they get better.”

Such statements “eviscerate” the AI companies’ fair use claim, the news agencies’ attorneys said.

Ripped From Today’s Headlines

The New York Times isn’t the only news outlet challenging AI firms and their fair use claim.

News Corporation, the parent company of the New York Post and The Wall Street Journal, is suing an AI startup that boasts it allows users to “skip the links” by providing them information without having to go through media sites.

The company, Perplexity AI, said it is “the world’s first ‘answer engine’—an easy-to-use tool that allows the public to obtain up-to-date, concise, factual answers to questions.”

To do so, it accesses copyrighted material, but also provides links to that material—if the customer wants them.

But if customers skip the links, that eats into the news sources’ profits, according to the suit.

“This conduct by Perplexity, from wherever it occurs, is intended to and/or foreseeably does in fact deprive [the plaintiffs] of customers, advertising revenues, and subscription revenues,” their complaint says.

Perplexity responded by saying that its AI outputs are not pre-fabricated, they are unpredictable and based off customer queries.

“The fact that, after an undisclosed number of attempts, Plaintiffs’ employee allegedly managed to coax Perplexity to provide answers that Plaintiffs claim are infringing says nothing about whether any other user, in New York or elsewhere, has ever sought or received the same or similar answers,” their court filing says. link

That case is ongoing.

Getty Images has also brought suit in the UK, alleging that startup Stability AI used its massive stockpile of photos to train its model, Stable Diffusion. The model was able to output distorted versions of images from Getty’s library—and could add a fake Getty watermark.

Stability AI argued that it wasn’t responsible for a third party using its product to add the watermark.

According to British High Court Justice Joanna Smith, “there is no evidence of a single user in the UK generating either Getty Images” or the watermarks using Stability’s later model, although it may have occurred in earlier, open-source versions.

Ultimately, the court ruled that it did not have jurisdiction for the infringements because there was no evidence they took place inside the UK.

Shredding Books

In July, a federal judge approved a $1.5 billion settlement between Anthropic and a group of plaintiff authors who alleged the company used pirated copies of their work—illegally downloaded from the internet—to train its model, Claude.

The company was originally on the hook for tens of billions of dollars; the case involved more than 480,000 books, each one potentially carrying a penalty of up to $150,000 under U.S. copyright law.

In that case, the judge said using the books to train AI models was transformative fair use; the violation was in using the pirated copies.

Another, similar suit against chipmaker Nvidia is ongoing. Although five authors filed suit in 2024, the case is still in the discovery phase, and the court is collecting evidence.

In a court filing, Nvidia cited fair use and denied any copyright infringement, but admitted it used “a third-party dataset referred to as ‘The Pile’” to train its model Nemo Megatron.

The authors in the case allege that The Pile—an 800 gigabyte text file compiled by a research company to train AI models—contained another dataset, Books3, which had copies of their work.

Books3 was originally hosted by AI startup Hugging Face, but was taken down for “reported copyright infringement” in 2023, their complaint says.

Song Lyrics

A band of music publishers is also suing Anthropic, alleging users are illegally accessing song lyrics through its Claude model. Although there are websites users can access to look up song lyrics, those sites have to pay a licensing fee.

The plaintiffs, led by Concord Music Group, also said Claude is outputting new songs based on chord progressions and nearly identical lyrics of copyrighted works, even when users didn’t specifically ask for them.

“For example, when asked by Publishers’ investigator to write ‘a song about the death of Buddy Holly,’ Claude generated a verbatim copy of the lyrics to ‘American Pie,’” the complaint said.

Meanwhile, Anthropic said Claude has “copyright-protective” features and that the company cannot be held responsible when a user finds a workaround.

“When a system ‘creates the copy only in response to the user’s command,’ its operator does not engage in volitional conduct,” Anthropic said in a court filing. “The user does.”