Publisher Elsevier's AI (artificial intelligence) on the first day of the London Book Fair held at the Olympia London exhibition space on March 11, 2025 in London, UK. London Book Fair is the annual international book industry marketplace for the negotiation of rights, sale and distribution of published content across print, audio, television, film and digital channels. 'LBF' typically attracts 25,000 visitors and exhibitors during the three-day event. (Photo: Richard Baker / In Pictures via Getty Images)
Photo from Getty Images
2025 has been a landmark year for litigation over content used to train LLMs and other generative AI systems, with creators benefiting from litigation in various courts and IP owners escalating legal actions against platforms for infringement on various fronts. But experts warn that the string of victories may be short-lived due to changes in the way LLMs are trained.
“Creators need to come to some kind of compensation agreement with AI companies soon, or they might get nothing,” Grammy-nominated electronic musician and producer BT warned during a panel discussion on AI and copyright at this year's Consumer Electronics Show.
He and other panelists were discussing the changing legal landscape as content owners are making stronger claims against technology companies for using their data without consent. Since the introduction of AI tools like OpenAI's ChatGPT and Google Gemini, controversy has arisen over how data is obtained, often indiscriminately harvesting content from the internet, regardless of whether the content is copyrighted or not.
In 2025, Anthropic, the developer of LLM Claude, agreed to a massive $1.5 billion settlement with the author after evidence showed that the company intentionally trained its models on millions of copyrighted works. Despite the settlement amount awaiting final court approval, many legal observers believe Anthropic avoided even larger damages based on established precedent.
In June, Meta won a narrow victory in its AI lawsuit, but court records revealed some damning evidence about the company's practices, and a judge left the door open to future lawsuits on stronger grounds. Additionally, Warner Bros. and other owners of global blue-chip IP have filed suit against Midjourney, alleging that they trained image-generating AI models on proprietary content, in addition to intellectual property claims against Stable Diffusion, Deviant Art, and others, proving that the output of those products closely resembled commercial property.
These successes have helped major content owners such as Condé Nast, Reuters, Shutterstock, and the New York Times secure deals with big tech companies and secure rights and compensation.
“Using our content without consent is neither appropriate nor productive,” Rebecca Grossman Cohen said. new york times At the CES panel. “Misuse must be rewarded.”
Grossman Cohen discussed some of the terms negotiated by the Times. This included fees for the use of the paper's extensive archives to train LLMs, as well as usage guidelines on how real-time information from the Times should be summarized into AI-generated search results.
Dave Davis, chief content officer of Protege, a new company formed to license content to AI companies under financial terms that are fair to creators, said the company's revenue will grow from zero to “eight digits” in 2025 as tech companies line up to secure access to rich, annotated, multimodal data (video, audio, text) through legitimate channels. This suggests that legal challenges could change the “break and grab” mentality that characterized the first phase of the AI data gold rush, opening the door to more lucrative deals for proactive creators.
Despite this string of successes, creators like BT warn that content owners need to act quickly to secure all kinds of terms. “A lot of artists struggle with AI,” he says. “In fact, if we don't reach some kind of agreement, we may end up with nothing.”
Of concern is that AI models are increasingly being trained on synthetic data, i.e., the output of AI systems rather than content attributed to individual creators or rights holders. Gartner predicts that 75% of AI training data in 2026 will be synthetic. This number could reach 100% by 2030. Once technology companies no longer need human-created content, they will stop paying for it.
“The quality of output from AI systems is improving dramatically, which means they can be trained on synthetic data without risking model collapse,” said Dr. Daniela Braga, founder and CEO of data training company Defined.ai, in a separate interview at CES. “The window of opportunity for individual rights holders to obtain favorable terms is steadily closing.”
Braga said the best way for creators to protect themselves is to do business with ethical companies that are willing to pay for high-quality, human-produced content and demonstrate superior value for that content to their customers. As model capabilities grow, the need shifts from vast amounts of data to data that is properly tagged and annotated to easily fit specific use cases.
Demand for services among businesses and consumers has lagged behind the massive and costly addition of capacity, and some serious questions remain about the sustainability of AI from a business perspective. For some artists who oppose generative AI itself, there may be a temptation to wait until the bubble bursts. After all, these artists created works for human enjoyment, not for mass consumption by machines that threaten our lives. Given these objections, the prospect of a meager payout may seem unattractive.
But if the industry's trajectory comes even remotely close to what AI advocates predict, creator advocates recommend taking the money while it's still on the table. Because the era of unwilling and even disastrous reconciliation may come to an end sooner than anyone expects.
