NYT Trial: Microsoft and OpenAI Warned of a "Vicious Cycle"

Key Takeaways

1Court documents show that Microsoft and OpenAI internally described a "vicious cycle" where LLMs undermine the web and their own content supply chain.

2The documents mention ChatGPT's memorization of protected content, the lack of systematic filtering of paid content, and examples of reproducing texts from the press.

3Experts at OpenAI estimate that AI-generated summaries may have reduced referral traffic to certain sites by 60%.

Why it matters : These revelations illustrate the awareness among major AI players of the risks their models pose to the information ecosystem and their own data sources.

Documents from the lawsuit filed by the New York Times reveal internal warnings at Microsoft and OpenAI: a "vicious cycle" between giant models and the web, with effects on model performance and the economics of websites. The documents also mention the memorization of protected content, the lack of filtering for paywalls, and a potential drop in referral traffic. Microsoft distances itself from the comments of one of its executives and insists on the consistency of its position in court.

Models Accused of Eroding Their Own Content Supply

Microsoft describes the content strategy surrounding LLMs as the trigger for a "vicious cycle" that harms both model performance and the web. This same document highlights that it is unusual for a final product to threaten the economic foundations of its essential suppliers, which would be the case for LLM activities and their "content supply chain." Microsoft has also admitted that LLMs represent "a product that destroys its own supply chain." According to internal documents from OpenAI, ChatGPT is presented as a "modern newsstand," while Nick Turley indicates that once the chatbot's response is obtained, there is "no good reason to click" on the source. Satya Nadella has stated that chatbots have replaced search and eliminated the need to access the original source. According to media and economics specialists at OpenAI, the reduction in referral traffic for sites like the New York Times is linked to AI-generated summaries, notably Google’s AI Overviews, and they estimate this drop in search-derived referrals at 60%.

Paywalled Data, Memorization, and Reproduction Risks

Satya Nadella has been quoted in favor of the principle that "everything that is paid should be licensed." An OpenAI representative admitted to being unaware of initiatives aimed at detecting or removing data from paywalled content from training corpora. Internal documents from OpenAI indicate that the company was aware that ChatGPT tended to reproduce passages "verbatim" from protected content and emphasized the importance of preventing memorization to limit copyright violations. Employees acknowledged that GPT-4 had memorized a significant amount of data and would be "incredibly good at regurgitating." ChatGPT has generated long excerpts from articles from various news outlets, such as the New York Times, Mercury News, Denver Post, LifeHacker, and Eurogamer. Microsoft also recognized that most authors did not anticipate this type of use of their work and were not compensated for it.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Microsoft Defends Its Stance and Isolates Brent Hecht's Comments

Microsoft has distanced itself from statements attributed to Brent Hecht. Its spokesperson Alex Haurek stated that these comments reflect an individual perspective, do not constitute legal analysis, and do not represent the company's position. In a court document, Jordan Usdan described Brent Hecht's role as "adversarial" and claimed that he is employed to provide divergent, academic, and futuristic viewpoints, without speaking on behalf of Microsoft on these theoretical issues. Alex Haurek also argued that Satya Nadella's testimony and Microsoft's position in the proceedings are consistent, noting that Nadella was discussing general principles and evolving uses, distinct from the legal conclusions addressed in the company's filings.

Internal Alerts and Statements on Impact for Publishers

The introduction of the case cites Brent Hecht and a ChatGPT official from OpenAI in connection with the idea of an "existential threat" to publishers such as the New York Times. Brent Hecht describes the data collection by ChatGPT and Copilot as "the greatest theft of labor in human history" and believes that the "fair use" argument makes it a "complete mockery." Statements from Satya Nadella, Sam Altman, and other OpenAI employees are also reported. Corporate alerts mention a "vicious cycle" harmful to the web and very critical judgments regarding how data is collected and legally defended.

Financial Goals Mentioned at OpenAI

Greg Brockman referred to the "gazillions" of dollars potentially generated by commercial AI. Jack Clark, then director of policy at OpenAI, described the creation of systems that replace the work of people who shape the "culture" of society.