Welcome to Eye on AI! In this edition…The New York Times and Amazon announce AI licensing deal…Elon Musk tried to block Sam Altman’s Middle East AI deal…Anthropic CEO warns AI could wipe out half of all entry-level white-collar jobs.
The daily drumbeat of AI hype cycles and tech breakthroughs often feels like riding big waves. But questions about the copyrighted data used to train generative AI models without owner consent have simmered for years—since before ChatGPT burst onto the scene in 2022.
Much of the legal wranglings have focused on the legal doctrine of “fair use”—that you can use someone else’s copyrighted work without permission—as long as the use is limited, transformative, and doesn’t hurt the original creator’s market.
One example is the ongoing case against Meta and OpenAI by comedian Sarah Silverman and two other authors, who filed copyright infringement lawsuits in 2023 alleging that pirated versions of their works were used without permission to train AI language models. The defendants recently argued that the use falls under fair use doctrine because AI systems “study” works to “learn” and create new, transformative content.
Federal district judge Vince Chhabria pointed out that even if this is true, the AI systems are “dramatically changing, you might even say obliterating, the market for that person’s work.” But he also took issue with the plaintiffs, saying that their lawyers had not provided enough evidence of potential market impacts.
Such is the convoluted logic puzzle around AI and copyright that it seems no one can fully solve—where both sides have compelling, seemingly reasonable arguments, yet the law doesn’t quite know how to reconcile them. Twentieth century copyright rules were designed for human creators, not 21st century AI systems trained on millions of works at once.
A seemingly logical argument
I recently came upon what seemed like a simple, logical argument that highlights the contradiction of applying copyright rules to generative AI. David Atkinson, a lecturer on law, ethics and AI at the University of Texas at Austin, asked on LinkedIn: “Serious question: Why should my students have to pay to use a textbook rather than be allowed to download from pirate sites, but $60 billion Anthropic shouldn’t?”
Atkinson was referring to another ongoing legal case filed against Anthropic by three authors, who alleged that Anthropic infringed upon their copyrights by using pirated versions of their books to train its AI systems. Anthropic’s argument is that as long as Anthropic is not outputting plagiarized content, then the actual training of the model should be fair use of the material.
But learning, Atkinson explained, is a transformative use—so why can’t his students argue that they’re taking books from pirated sites in order to learn something?
I loved the idea of such a tidy, logical notion. If one is fair use, why isn’t the other? But when I reached out to several lawyers on the issue, I received several thoughtful rebuttals that left me as uncertain as ever.
Not quite apples to apples
While Atkinson’s argument is “very clever,” it “interprets AI as essentially eviscerating all IP law,” Bradford Newman, who leads the AI and blockchain practice at law firm Baker McKenzie, told me. But from an intellectual property perspective, there is a big difference between the AI training and a student downloading pirated content, he explained.
If someone wants to download a specific book or watch a specific movie and downloads it for free, that’s clear copyright infringement, he said. But that’s harder to address with an AI model.
“If you trained an algorithm on one copyrighted work only, that would be a problem,” he said, but market harm in AI training is harder to prove. “Training on many books doesn’t necessarily harm any specific author’s market—especially if the model isn’t reproducing their work verbatim.”
Katie Gardner, a partner at Gunderson Dettmer, agreed that the student piracy analogy “oversimplifies” how copyright law works. “When a student pirates a textbook, that’s classic market substitution since they’re avoiding a purchase they would have otherwise made,” she explained. AI training is essentially teaching machines to understand language patterns, she added, and early indications from some courts suggest that using copyrighted works to train AI models may qualify as sufficiently transformative fair use.
The real legal flashpoint, she added, isn’t the training process—it’s the output. “If an AI system can generate content that substitutes for the original copyrighted work, that’s where publishers/authors have their strongest case,” she said. “Courts tasked with developing the new rules will focus on this market reality: Does the AI output substitute for the original work or create something genuinely new?”
A case of unsettled law
The twisted pretzel of logic, however, is clearly why the U.S. Copyright Office weighed in on the arguments in early May, with a 108-page report on whether the unauthorized use of copyrighted materials to train generative AI systems is fair use.
The answer? It’s complicated.
The rules of fair use, the copyright office said, depends on context—how the AI is used and what it produces. Training AI on copyrighted books, songs, or images might be fair use in some situations, but it’s not automatically allowed. If the AI creates content that’s similar in style to the original (like writing a poem that sounds like a famous poet or mimicking a best-selling novel), then it’s probably not very transformative—a key requirement for fair use.
On the other hand, the copyright office emphasized that AI training is not fair use because it’s “just learning” like a human. AI does not forget its trained content or “reinterpret” it, as humans do. It also pointed out that if the AI produces outputs that compete in the same market, even if the content isn’t identical, it could still hurt the original creators and tip the scale against fair use. But guardrails, like blocking prompts that produce infringing outputs, could support a fair-use argument.
Does your head hurt yet? Mine does. The bottom line is that the law remains unsettled around fair use. Newman has long argued that these issues could make their way to the Supreme Court. For now, we’ll have to continue to let them simmer.
With that, here’s the rest of the AI news.
Sharon Goldman
sharon.goldman@fortune.com
@sharongoldman
AI IN THE NEWS
New York Times and Amazon announce AI licensing deal. According to the New York Times, the company has agreed to license its editorial content to Amazon for use in the tech giant’s artificial intelligence platforms. The agreement includes not only news articles, but also content from NYT Cooking, the Times’s popular recipe site, and The Athletic, its sports-focused publication. This marks the first time the Times has entered into a licensing agreement specifically tied to generative AI technology. The deal comes against the backdrop of an ongoing legal battle: In 2023, the Times filed a copyright infringement lawsuit against OpenAI and Microsoft, alleging the companies used millions of its articles to train AI chatbots without permission or compensation. Both OpenAI and Microsoft have denied the claims.
Elon Musk tried to block Sam Altman’s Middle East AI deal. Elon Musk’s work with the Trump Administration is already in the rear-view mirror, but before announcing he was leaving DOGE he unsuccessfully tried to block Sam Altman’s Middle East AI deal, according to the Wall Street Journal. After OpenAI won a deal to build one for the world's largest AI data centers in Abu Dhabi, an expansion of its Stargate project and the first of its “OpenAI for Countries” initiatives, Musk worked hard behind the scenes to derail the deal if it did not include his startup xAI.
Anthropic CEO warns AI could wipe out half of all entry-level white-collar jobs. Anthropic CEO Dario Amodei told Axios that he believes AI could wipe out half of all entry-level white-collar jobs—and spike unemployment to 10-20% in the next one to five years. Amodei said AI companies and governments “need to stop ‘sugar-coating’ what’s coming: the possible mass elimination of jobs across technology, finance, law, consulting and other white-collar professions, especially entry-level gigs.” The interview comes at a moment when companies like Shopify, Duolingo, and Fiverr have warned of AI-based layoffs.
FORTUNE ON AI
Nvidia’s CEO blasted Trump policy that will cost the company $10.5 billion in lost revenue—then praised Trump’s ‘bold vision’ minutes later —by Alexandra Sternlicht
OpenAI wants to help countries develop their own AI capabilities. But can they afford it? —by Leo Schwartz
Malaysia’s prime minister thinks AI may need to operate independently of national sovereignty —by Nicholas Gordon
‘Boil the ocean’ vs. ‘fry the fish’: How smaller countries can carve out an AI niche between the superpowers —by Leo Schwartz
AI CALENDAR
June 9-13: WWDC, Cupertino, Calif.
July 13-19: International Conference on Machine Learning (ICML), Vancouver
July 22-23: Fortune Brainstorm AI Singapore. Apply to attend here.
Sept. 8-10: Fortune Brainstorm Tech, Park City, Utah. Apply to attend here.
Oct. 6-10: World AI Week, Amsterdam
Dec. 2-7: NeurIPS, San Diego
EYE ON AI NUMBERS
20%
AI now accounts for as much as 20% of global data center electricity use—and that figure could nearly double by the end of this year, according to new research published Thursday in Joule and reported by Wired. The analysis, authored by Alex de Vries-Gao, founder of Digiconomist, warns that AI’s energy consumption is on track to surpass that of Bitcoin mining by the end of 2025. De Vries-Gao, who launched Digiconomist in the late 2010s to track the environmental impact of crypto, says the urgency around AI’s power demands has surged with the explosive growth of energy-hungry large language models like ChatGPT.