Aura Kasih
Newly disclosed legal documents, unsealed as part of The New York Times’ lawsuit against OpenAI and Microsoft, provide a stark glimpse into the internal deliberations and concerns of these AI giants regarding the foundational impact of their generative AI technologies on the internet and its content ecosystem. The filings suggest a profound awareness within both organizations that their data-scraping practices and the subsequent AI models trained on this data were initiating a detrimental cycle, threatening the very fabric of online information creation and consumption.
The unsealed documents present a series of internal communications and testimonies that paint a picture of companies grappling with the ethical and economic ramifications of their AI development. Far from being unaware of potential harms, evidence indicates that key figures within OpenAI and Microsoft foresaw and, in some instances, explicitly articulated the negative consequences of their approach. These revelations challenge the narrative of nascent technology blindly stumbling into unforeseen problems, instead pointing to a calculated progression with acknowledged, albeit perhaps downplayed, risks.
Central to the revelations are the candid remarks attributed to Brent Hecht, identified as Microsoft’s Director of Applied Science. His documented assessments are particularly striking, characterizing the process of data acquisition for AI training as "an astonishing theft of unprecedented proportions" and, more damningly, as potentially "the largest theft of labor in human history." These statements, made internally, directly confront the notion of "fair use" that has been a cornerstone of digital copyright law. Hecht’s assertion that Microsoft’s defense of its practices would "make a complete mockery of the idea of ‘fair use’" underscores the internal recognition of a significant legal and ethical challenge.
Microsoft, in its legal defense, has attempted to distance itself from these specific statements, with spokesperson Alex Haurek stating they represent "one employee’s individual perspective" and "do not represent the company’s views." Further efforts to contextualize Hecht’s role were made by Jordan Usdan, GM for Data Strategy and Ops at Microsoft AI, who described Hecht as holding "divergent, academic, and forward-looking views" and not as someone who speaks for the company on theoretical impacts to content creators. However, the sheer volume and content of these internal documents suggest that such perspectives were not isolated but rather part of a broader internal discourse, even if not officially sanctioned as company policy.
The core of the concern lies in what internal documents have termed a "doom loop." This concept describes a self-perpetuating cycle where AI models, trained on vast amounts of web content, begin to replace the need for users to visit the original sources. This erosion of traffic and revenue for content creators, in turn, diminishes the very pool of high-quality, original content available for future AI training, potentially leading to a degradation of AI model performance and the overall health of the internet. Microsoft CEO Satya Nadella himself acknowledged under oath that interacting with chatbots "has substituted…giving you the information right there on the website on the AI platform versus needing to go to the underlying source." An internal Microsoft document explicitly articulated this danger: "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.’"
Beyond the existential threat to the web’s content ecosystem, the documents also reveal a stark financial motivation behind the AI development. OpenAI co-founder Greg Brockman is quoted as being "deeply motivated by the gazillions" he hoped to gain by commercializing OpenAI’s technology. This financial ambition, juxtaposed with the potential damage to content creators, raises critical questions about the ethical underpinnings of the pursuit of artificial general intelligence. The prospect of an OpenAI IPO, reportedly based on a trillion-dollar valuation, further highlights the immense commercial stakes involved.
The methods of data acquisition have also come under scrutiny. Despite later statements from Nadella emphasizing that "anything that is paywalled should be licensed," an OpenAI representative admitted to being "unaware" of any efforts to detect or remove paywalled content from their training datasets. This suggests a practice of circumventing access controls and terms of use, effectively treating copyrighted material as freely available data. The legal implications of such actions, particularly concerning the doctrine of fair use, are a central tenet of the ongoing lawsuit.
Furthermore, internal communications within OpenAI reveal a clear understanding of the potential for their models to simply reproduce copyrighted material verbatim. As early as 2020, OpenAI recognized that its API "might output existing content verbatim." By 2021, the company deemed the "prevention of memorization" crucial for "fair use [compliance] and minimizing copyright violations in model output." Yet, by June 2022, employees acknowledged that GPT-4 would have "memorized a ton of data and therefore will be insanely good at regurgitation." The lawsuit cites examples where ChatGPT has reproduced substantial portions of articles from various publications, including The Times, directly in response to user queries.
This "regurgitation" capability is directly linked to the "doom loop" phenomenon. When AI models can directly answer user queries with content previously published elsewhere, the incentive for users to visit the original source diminishes significantly. This has a direct impact on referral traffic, a critical revenue stream for publishers. OpenAI’s own economic experts, such as Dr. Goldfarb, have opined that declines in referral traffic to publications like The Times are "driven by a combination" of factors, including "e.g., Google AI Overviews." These AI-driven summaries, by providing immediate answers, can depress search referrals by significant margins, as suggested by estimates ranging from 20 to 60 percent for some publications. Dr. Sinnreich, an OpenAI media expert, further supported this by citing an analysis showing substantial drops in referral traffic from both Google Search and Google Discover since the introduction of AI overviews.
The very nature of large language models (LLMs) is perceived by some within the companies as inherently disruptive to their "content supply chain." Microsoft itself acknowledged that "LLMs are a product that destroys its own supply chain." This is because the models become substitutes for the very information work that generated the training data. This substitution directly impacts the "labor of the people" who create content, including journalists, writers, and artists. The documents highlight a "real risk" that generative AI could "significantly disrupt[] the employment of the very people who generated the data on which the foundation model was trained."
In essence, the unsealed documents suggest that OpenAI and Microsoft were not merely passive observers of the disruptive potential of their technologies. Instead, they appear to have possessed a detailed, albeit often internally expressed, understanding of the profound and potentially irreversible damage their AI models could inflict upon the web’s content ecosystem, its creators, and the economic models that sustain them. The pursuit of "gazillions" in potential profit, as one co-founder candidly put it, seems to have outweighed the acknowledged risks of initiating a "doom loop" that could ultimately undermine the internet as a vibrant platform for information and creativity.
Microsoft’s spokesperson reiterated that Nadella’s testimony and the company’s position are "perfectly consistent," focusing on broad principles of information consumption rather than specific copyright conclusions. However, the cumulative weight of the evidence presented in these unsealed documents paints a compelling picture of companies forging ahead with technologies whose detrimental effects on the web and its creators were not only foreseen but, in many instances, explicitly documented and discussed internally. The ongoing legal battle will likely hinge on how these internal acknowledgments are weighed against the companies’ public stances and legal defenses concerning copyright and fair use in the age of artificial intelligence. The implications extend far beyond this specific lawsuit, touching upon the future of content creation, intellectual property, and the fundamental architecture of the internet itself.
Safa Marwah
Ayu Aulia
