Artificial intelligence companies have long argued that the web provides the knowledge needed to build increasingly capable AI models. But according to newly surfaced court documents, Microsoft employees privately acknowledged a growing problem: AI-generated content could end up poisoning the very information ecosystem future AI systems rely on.

The documents, highlighted in reporting by Futurism and 404 Media, were filed as part of the ongoing copyright lawsuit against OpenAI and Microsoft. They include internal Microsoft discussions describing an AI “doom loop” in which AI-generated content gradually overwhelms human-created material online, making it increasingly difficult to train future large language models on high-quality data.

Microsoft reportedly warned about its own “content supply chain”

One of the most striking passages comes from an internal Microsoft document discussing what employees referred to as the company’s “content supply chain.”

According to the filing, Microsoft acknowledged that the widespread generation of AI-written articles, websites, and other online content could eventually undermine the economic incentives for publishers and creators to produce original work. The document reportedly states that it is highly unusual for a product to threaten “the economic foundations of its essential suppliers”—yet argues that this is exactly what generative AI risks doing to the web that supplies its training data.

The concern isn’t simply about quality. If fewer people can make a living producing original journalism, books, tutorials, reviews, research, and other human-created material, there is less fresh information available for future AI models to learn from.

The “AI slop” problem

Researchers and critics have increasingly warned about a feedback loop where AI systems train on content that was itself produced by earlier AI models.

This phenomenon is sometimes described as model collapse or a “doom loop.” As more AI-generated text floods search results, blogs, and social media, distinguishing original reporting from synthetic content becomes increasingly difficult. Training future models on this recycled material could gradually reduce accuracy, originality, and factual reliability.

Microsoft executive Ryan Roslansky has also publicly warned about AI-generated workplace documents creating their own “doom loop,” where employees repeatedly summarize AI-written summaries rather than producing original work.

Copyright arguments are at the center of the lawsuit

The documents emerged during ongoing copyright litigation brought by news organizations against OpenAI and Microsoft.

According to 404 Media‘s reporting, plaintiffs argue the internal communications undermine one of the companies’ legal defenses by showing employees understood that AI systems depend on copyrighted human-created works while simultaneously threatening the businesses that produce them. The filings reportedly also reference internal acknowledgements that large language models were trained using vast amounts of online content without individual licensing agreements.

Microsoft and OpenAI have consistently argued in court that training AI models on publicly available material constitutes fair use under U.S. copyright law, while many publishers and creators disagree. Multiple lawsuits remain ongoing, and no final legal precedent has yet been established.

Why it matters

The newly surfaced documents don’t settle the legal questions surrounding AI training, but they provide a rare glimpse into how Microsoft employees reportedly discussed the long-term sustainability of generative AI behind closed doors.

If AI companies depend on a healthy ecosystem of human creators, publishers, and journalists for future training data, then protecting that ecosystem may become more than a copyright issue—it could become essential to the future quality of AI itself.