Chatbot builder Anthropic agreed to pay $1.5 billion to authors in a landmark copyright settlement that could redefine how artificial intelligence companies compensate creators.

The San Francisco-based startup is ready to pay authors and publishers to settle a lawsuit that accused the company of illegally using their work to train its chatbot.

Anthropic developed an AI assistant named Claude that can generate text, images, code and more. Writers, artists and other creative professionals have raised concerns that Anthropic and other tech companies are using their work to train their AI systems without their permission and not fairly compensating them.

As part of the settlement, which the judge still needs to be approve, Anthropic agreed to pay authors $3,000 per work for an estimated 500,000 books. It’s the largest settlement known for a copyright case, signaling to other tech companies facing copyright infringement allegations that they might have to pay rights holders eventually as well.

Meta and OpenAI, the maker of ChatGPT, have also been sued over alleged copyright infringement. Walt Disney Co. and Universal Pictures have sued AI company Midjourney, which the studios allege trained its image generation models on their copyrighted materials.

“It will provide meaningful compensation for each class work and sets a precedent requiring AI companies to pay copyright owners,” said Justin Nelson, a lawyer for the authors, in a statement. “This settlement sends a powerful message to AI companies and creators alike that taking copyrighted works from these pirate websites is wrong.”

Last year, authors Andrea Bartz, Charles Graeber and Kirk Wallace Johnson sued Anthropic, alleging that the company committed “large-scale theft” and trained its chatbot on pirated copies of copyrighted books.

U.S. District Judge William Alsup of San Francisco ruled in June that Anthropic’s use of the books to train the AI models constituted “fair use,” so it wasn’t illegal. But the judge also ruled that the startup had improperly downloaded millions of books through online libraries.

Fair use is a legal doctrine in U.S. copyright law that allows for the limited use of copyrighted materials without permission in certain cases, such as teaching, criticism and news reporting. AI companies have pointed to that doctrine as a defense when sued over alleged copyright violations.

Anthropic, founded by former OpenAI employees and backed by Amazon, pirated at least 7 million books from Books3, Library Genesis and Pirate Library Mirror, online libraries containing unauthorized copies of copyrighted books, to train its software, according to the judge.

It also bought millions of print copies in bulk and stripped the books’ bindings, cut their pages and scanned them into digital and machine-readable forms, which Alsup found to be in the bounds of fair use, according to the judge’s ruling.

In a subsequent order, Alsup pointed to potential damages for the copyright owners of books downloaded from the shadow libraries LibGen and PiLiMi by Anthropic.

Although the award was massive and unprecedented, it could have been much worse, according to some calculations. If Anthropic were charged a maximum penalty for each of the millions of works it used to train its AI, the bill could have been more than $1 trillion, some calculations suggest.

Anthropic disagreed with the ruling and didn’t admit wrongdoing.

“Today’s settlement, if approved, will resolve the plaintiffs’ remaining legacy claims,” said Aparna Sridhar, deputy general counsel for Anthropic, in a statement. “We remain committed to developing safe AI systems that help people and organizations extend their capabilities, advance scientific discovery, and solve complex problems.”

The Anthropic dispute with authors is one of many cases where artists and other content creators are challenging the companies behind generative AI to compensate for the use of online content to train their AI systems.

Training involves feeding enormous quantities of data — including social media posts, photos, music, computer code, video and more — to train AI bots to discern patterns of language, images, sound and conversation that they can mimic.

Some tech companies have prevailed in copyright lawsuits filed against them.

In June, a judge dismissed a lawsuit authors filed against Facebook parent company Meta, which also developed an AI assistant, alleging that the company stole their work to train its AI systems. U.S. District Judge Vince Chhabria noted that the lawsuit was tossed because the plaintiffs “made the wrong arguments,” but the ruling didn’t “stand for the proposition that Meta’s use of copyrighted materials to train its language models is lawful.”

Trade groups representing publishers praised the Anthropic settlement on Friday, noting it sends a big signal to tech companies that are developing powerful artificial intelligence tools.

“Beyond the monetary terms, the proposed settlement provides enormous value in sending the message that Artificial Intelligence companies cannot unlawfully acquire content from shadow libraries or other pirate sources as the building blocks for their models,” said Maria Pallante, president and chief executive of the Association of American Publishers in a statement.

The Associated Press contributed to this report.

OpenAI researchers claim they’ve cracked one of the biggest obstacles to large language model performance — hallucinations.

Hallucinations occur when a large language model generates inaccurate information that it presents as fact. They plague the most popular LLMs, from OpenAI’s GPT-5 to Anthropic’s Claude.

OpenAI’s baseline finding, which it made public in a paper released on Thursday, is that large language models hallucinate because the methods they’re trained under reward guessing more than admitting uncertainty.

In other words, LLMs are being told to fake it till they make it. Some are better than others, however. In a blog post last month, OpenAI said that Claude models are more “aware of their uncertainty and often avoid making statements that are inaccurate.” It also noted that Claude’s high refusal rates risked limiting its utility.

“Hallucinations persist due to the way most evaluations are graded — language models are optimized to be good test-takers, and guessing when uncertain improves test performance,” the researchers wrote in the paper.

Large language models are essentially always in “test-taking mode,” answering questions as if everything in life were binary — right or wrong, black or white.

In many ways, they’re not equipped for the realities of life, where uncertainty is more common than certainty, and true accuracy is not a given.

“Humans learn the value of expressing uncertainty outside of school, in the school of hard knocks. On the other hand, language models are primarily evaluated using exams that penalize uncertainty,” the researchers wrote.

The good news is that there is a fix, and it has to do with redesigning evaluation metrics.

“The root problem is the abundance of evaluations that are not aligned,” they wrote. “The numerous primary evaluations must be adjusted to stop penalizing abstentions when uncertain.”

In a blog post about the paper, OpenAI elaborated on what this type of adjustment would entail.

“The widely used, accuracy-based evals need to be updated so that their scoring discourages guessing. If the main scoreboards keep rewarding lucky guesses, models will keep learning to guess,” OpenAI said.

OpenAI did not immediately respond to a request for comment from Business Insider.

“It’s not true, and in fact, it’s kind of the opposite,” Karp said in an interview with Fortune Thursday at the company’s commercial customer conference, AIPCon, where Palantir customers showcased how they were using the company’s software platform and generative AI within their own businesses at George Lucas’ Skywalker Ranch in Marin County, Calif. 

The primary danger of AI in this country, says Karp, is that workers don’t understand that AI will actually help them in their roles—and it will hardly replace them. “Silicon Valley’s done an immensely crappy job of explaining that,” he said. “If you’re in manufacturing, in any capacity: You’re on the assembly line, you maintain a complicated machine—you have any kind of skilled labor job—the way we do AI will actually make your job more valuable and make you more valuable. But currently you would think—just roaming around the country, and if you listen to the AI narratives coming out of Silicon Valley—that all these people are going to lose their jobs tomorrow.”

Karp made these comments the day before the Bureau of Labor Statistics released its August jobs report, which showcased a climbing unemployment rate and stagnating hiring figures, reigniting fears of whether AI is at all responsible for the broader slowdown. There has been limited data thus far suggesting that generative AI is to blame for the slowing jobs market—or even job cuts for that matter—though a recent ADP hiring report offered a rare suggestion that AI may be one of several factors influencing hiring sentiment. Some executives, including Salesforce’s Marc Benioff, have cited the efficiency gains of AI for layoffs at their companies, and others, like Ford CEO Jim Farley and Amazon CEO Andy Jassy, have made lofty predictions about how AI is on track to replace jobs in the future. Most of these projections have been centered around white collar roles, in particular, versus manufacturing or skilled labor positions.

Karp, who has a PhD in neoclassical social theory and a reputation for being outspoken and contrarian on many issues, argues that fears of AI eliminating skilled labor jobs are unfounded—and he’s committed to “correcting” the public perception. 

Earlier this week, Palantir launched “Working Intelligence: The AI Optimism Project,” a quasi-public information and marketing campaign centered around artificial intelligence in the workplace. The project has begun with a series of short blog posts featuring Palantir’s customers and their opinions on AI, as well as a “manifesto” that takes aim at both the “doomers” and “pacifiers” of AI. “Doomers fear, and pacifiers welcome, a future of conformity: a world in which AI flattens human difference. Silicon Valley is already selling such bland, dumbed-down slop,” the manifesto declares, arguing that the true power of AI is not to standardize but to “supercharge” workers.

Jordan Hirsch, who is spearheading the new project at Palantir, said that there are approximately 20 people working on it and that they plan to launch a corresponding podcast.

While Palantir has an obvious commercial interest in dispelling public fears about AI, Karp framed his commitment to the project as something important for society. Fears about job replacement will “feed a kind of weird populism based on a notion that’s not true—that’s going to make the factions on the right and left much, much, much more powerful based on something that’s not true,” he said. “I think correcting that—but not just by saying platitudes, but actually showing how this works, is one of the most important things we have to get on top of.”

Karp said he planned to invest “lots of energy and money” into the AI Optimism Project. When asked how much money, he said he didn’t know yet, but that “we have a lot of money, and it’s one of my biggest priorities.” 

Palantir has seen enormous growth within the commercial side of its business in the last two years, largely due to the artificial intelligence product it released in 2023, called “AIP.” Palantir’s revenue surpassed $1 billion for the first time last quarter. And while Palantir only joined the S&P 500 last year, it now ranks as one of the most valuable companies in the world thanks to its soaring stock price.