Court documents unsealed in the New York Times’ lawsuit against OpenAI and Microsoft reveal that the two companies internally knew their generative AI models threatened the web ecosystem they depend on for training data. Dubbed « the greatest heist of intellectual property in human history » by a Microsoft executive, the LLM ecosystem stands accused by its own architects of triggering a destructive spiral for media outlets.
🔑 Key takeaways
- An internal Microsoft memo describes a « doom loop » threatening LLMs and the web
- Microsoft’s Brent Hecht calls AI training « the greatest heist of intellectual property in human history »
- Click-through rates from Bing Chat to NYT sites drop by 87 % to 93 % vs traditional search
- CEO Satya Nadella admits chatbots have « replaced » visits to publisher sites
- The US Department of Justice files a statement of interest in favor of OpenAI and Microsoft
Microsoft’s « doom loop » memo
An internal Microsoft document, obtained by plaintiffs in the New York Times’ lawsuit, bluntly describes the company’s AI strategy as a threat to its own supply chain. « Our AI content strategy has started a doom loop that will simultaneously harm our models’ performance and the entire web, » reads the memo, revealed through court proceedings.
The document notes that it is « extremely unusual for a finished product to threaten the economic foundations of its essential suppliers, but that is exactly the situation we have created for our LLM business regarding its content supply chain. » In short, AI models are trained on content produced by publishers, then used to generate responses that deprive those same publishers of traffic and revenue — a vicious circle documented internally by Microsoft itself.
« Our AI content strategy has started a doom loop that will simultaneously harm the performance of our models and the entire web. »
Internal Microsoft document
Executives sounding the alarm
Several Microsoft and OpenAI leaders internally expressed concerns about their products’ impact. Brent Hecht, director of applied science at Microsoft, called the process of training on copyrighted content « the greatest heist of intellectual property in human history » and « an unprecedented heist of staggering scope. » He also described large language models as « a product that destroys its own supply chain. »
At OpenAI, Nick Turley, vice-president in charge of ChatGPT, warned that publishers face an « existential threat » from AI products, which he described as « largely substitutable » and that « will become increasingly substitutable as they improve. » An OpenAI software engineer testified that, regardless of the visibility given to source links, « users will not click. »
Microsoft CEO Satya Nadella himself testified under oath that chatbots have « replaced » visits to publisher sites by giving « the information directly on the AI platform rather than requiring the user to go to the source. » He added that any content behind a paywall should be licensed, and that he would have exercised the company’s right to demand model retraining had he known OpenAI had used paywalled content.
The scale of the phenomenon in numbers
Traffic data provided by Microsoft as part of the trial confirms publishers’ fears. Click-through rates from Bing Chat are significantly lower than those generated by traditional Bing searches, demonstrating a massive siphoning of traffic toward conversational interfaces.
| Source / Metric | Referral reduction |
|---|---|
| New York Times sites (Bing Chat vs Bing search) | -87 % to -93 % |
| New York Daily News sites (Bing Chat vs Bing search) | -83 % to -91 % |
| OpenAI estimate: overall drop in search referrals | up to -60 % |
Jack Clark, head of policy at OpenAI, further acknowledged that the company was building « systems that substitute for the work of the people who define society’s ‘culture.’ » Internal documents described ChatGPT as « the modern newsstand, » a formulation that captures the ongoing substitution.
Documented paywall circumvention
The file also reveals questionable internal practices at OpenAI. Greg Brockman, co-founder and president of the company, showed little concern for the legality of scraping. When an employee, Nick Ryder, flagged « a hack to bypass the New York Times paywall », Brockman simply replied « ah nice », according to court documents.
The case file also mentions custom GPTs designed to easily extract content from paywalled articles, with evocative names like « Bypass Paywall » and « Article Reader. » These revelations contradict OpenAI’s defense, which argues that training AI systems on copyrighted works falls under the fair use doctrine.
The DOJ sides with AI
The revelations come amid a sensitive legal backdrop. The US Department of Justice has filed a statement of interest in favor of OpenAI and Microsoft in this case, arguing that denying AI models access to copyrighted content « would impede creative and scientific progress while harming American prosperity and economic mobility. »
For its part, OpenAI has accused the New York Times of seeking « undue compensation at the expense of progress. » Both companies have tried to downplay the internal statements. A Microsoft spokesperson said « these comments reflect an individual employee’s perspective, do not constitute legal analysis, and do not represent the company’s positions. » OpenAI did not respond to requests for comment. The documents are part of a publishers’ motion for summary judgment — a court ruling without a full trial — and the court has not yet ruled on whether either company infringed publishers’ copyrights.
Conclusion: a ticking time bomb for the ecosystem
These revelations paint an unprecedented picture: the very architects of large language models have identified, documented, and acknowledged the existential risks they pose to the web ecosystem. The « doom loop » theorized in Microsoft’s memo is no longer a hypothesis: it is already measurable as a 87 % to 93 % drop in traffic for major publishers. In the long run, the disappearance of quality content sources could paradoxically degrade the models’ own performance, validating the memo’s prediction.
The New York Times’ lawsuit, along with several similar class actions, will determine whether the AI industry can continue to freely extract the value produced by media outlets, or whether a new licensing framework — like the ones already emerging from select partnerships — will eventually prevail. For publishers and investors exposed to the media sector alike, the stakes go beyond copyright: they concern the viability of the web’s open model.
Sources
This article is published for informational and educational purposes only. It does not constitute investment advice. Do your own research (DYOR) before making any decisions.
