AI 학습 '공정 이용' 논리, 내부 임원들조차 '경악할 도난'이라 불러 흔들린다

AI training built on fair use looks shaky when the companies' own people call it "astonishing theft" Matthias Bastian View the LinkedIn Profile of Matthias Bastian Sep 18, 2026 Nano Banana Pro prompted by THE DECODER Key Points Several US media companies, including The New York Times, are seeking billions of dollars in damages from OpenAI and Microsoft over alleged copyright infringement in AI training. The plaintiffs cite internal messages and sworn testimony in which executives questioned their own fair use defense and described chatbots as substitutes for original journalism. The filing also accuses OpenAI of systematically bypassing paywalls, violating license terms for training data, and deploying filters to suppress evidence after the lawsuits were filed. Ask about this article… Search A new court filing by The New York Times and other plaintiffs cites previously undisclosed internal emails and sworn testimony from OpenAI and Microsoft executives. The New York Times and several other media companies have filed a joint summary judgment brief in US District Court in New York. Joining the Times are the Daily News group, which includes the Chicago Tribune and Denver Post, and Ziff Davis , which owns CNET, IGN, and PCMag. The plaintiffs also include the Center for Investigative Reporting, home to Mother Jones, and The Intercept . The plaintiffs are seeking billions of dollars in damages, according to the Financial Times . The 92-page brief is part of consolidated multidistrict litigation that brings together several lawsuits, beginning with the New York Times suit filed in December 2023. Ad Executives' internal statements undercut the fair use defense The brief draws on internal emails, Slack messages, and sworn testimony disclosed during discovery, including statements that undercut the AI companies' fair use defense. Ad Microsoft's director of applied science, Brent Hecht, called the practice "an astonishing theft of unprecedented proportions" and possibly the "largest theft of labor in human history." He also wrote that a successful fair use defense would arguably "make a complete mockery of the idea of 'fair use.'" Microsoft told the Financial Times these comments "reflect one employee's individual perspective" and "are not a legal analysis." The US Copyright Office also concluded in May 2025 that fair use cannot apply broadly given the sheer scale at which AI companies copy data. The official who oversaw the report was later fired by the Trump administration . Ad Nick Turley, OpenAI's head of ChatGPT, wrote that publishers face an "existential threat," according to the brief. He said the products "are largely substitutive, period" and would increasingly replace publishers' offerings "as they get better." An OpenAI engineer added that "no matter how prominently we show the links, users won't click." Microsoft CEO Satya Nadella confirmed under oath that chatbot conversations had replaced visits to original sources. In a statement to the FT, Microsoft said his testimony concerned "broad principles and changes under way in how people find and consume information." It was not a "conclusion about copyright questions" at the center of the case, the company said. Ad An internal Microsoft document describes a self-reinforcing cycle it calls a "doom loop." Ad 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: It is highly unusual that an end-product threatens the economic foundations of its essential suppliers, but that is the situation we have created for our LLM business with respect to its 'content supply chain. Microsoft's own data shows that click-through rates on Copilot were much lower than on traditional Bing search. Rates were down 87 to 93 percent for the New York Times, 83 to 91 percent for the Daily News group, and 51 to 94 percent for Ziff Davis. OpenAI internally described local news, a core business for the Daily News group, as a "pretty common quer[y]" in ChatGPT. Other studies have also found that chatbot answers sharply reduce traffic to the open web . OpenAI co-founder Greg Brockman discussed the models' ability to handle news in an internal message. we are excellent at news btw. every time i do any generative stuff on NYT it seems to predict the next sentence pretty well. Paywall workarounds and license restrictions complicate OpenAI's defense The plaintiffs describe how OpenAI systematically bypassed paywalls and ignored terms of service. When an employee told Brockman about "a hack to get around nytimes paywall," he replied "ah nice." Around 2017, Brockman also wrote that he was "deeply motivated by the gazillions" he hoped to earn by commercializing OpenAI's technology. OpenAI's corporate representative testified that he knew of no method for detecting paywalled content in the training data and no effort to remove it. The company's standard crawling process "did not include reviewing websites['] ... Terms of Use or Service." Nadella testified under oath that "anything that is paywalled should be licensed by anyone who wants to use it." He said he would have forced OpenAI to retrain its models had he known the company had scraped paywalled content and used it for training. OpenAI also acquired the "New York Times Annotated Corpus," a collection of 1.8 million articles, through a third party. Its license restricted use to "non-commercial linguistic education, research and technology development." OpenAI employees knew using the corpus to train models "would not be appropriate," but did so anyway. Plaintiffs accuse OpenAI of suppressing evidence and challenge all four fair use factors Immediately after the lawsuits were filed, OpenAI built a filter to suppress output most likely drawn from the plaintiffs' publications, according to the brief. Content from companies that hadn't sued remained unaffected. The plaintiffs argue that the filter was designed not to protect copyrights but to prevent them from gathering evidence. Microsoft's Hecht had raised concerns that OpenAI might deploy such a filter, calling it an "accidental cover up." He warned it would result in "people who have a right over the content having less visibility into what was used for training." Even so, the brief includes numerous examples of problematic outputs, with ChatGPT reproducing exact copies and summaries of NYT articles on request, including articles behind a paywall. The plaintiffs also argue that the fair use defense fails on all four statutory factors. They say the use is substitutive and commercial rather than transformative, while the articles are expressive works at the core of copyright protection. The defendants also copied entire works, even though their own experts admitted that no single work was necessary. On market harm, the plaintiffs point to existing licensing markets and those that could realistically develop. OpenAI and Microsoft have signed licensing deals with other publishers, as have Amazon, Google, Meta, and Perplexity. But rather than pay for the plaintiffs' content, the defendants took it for free and traded it among themselves, the brief argues. A functioning licensing market makes the fair use argument harder to sustain. The defendants can't credibly claim licensing wasn't possible when they're already paying for comparable content from other publishers. AI can generate a million "pink slime" articles for $6,800 The plaintiffs also challenge the argument that AI models don't compete with publishers, further undercutting the fair use defense. They argue that OpenAI's products let users flood the market with "pink slime," meaning low-quality, often plagiarized pseudo-news. At OpenAI's API prices, generating one million news-style articles of 500 words each costs about $6,800, without involving a single journalist. The brief cites Prism News, which used just four employees to run 200 AI-generated publications posing as local newsrooms and hobby sites. Ope