Lawsuit Documents Detail Internal Microsoft and OpenAI Discussions on AI Data Practices

Documents filed in ongoing copyright litigation between news publishers and OpenAI and Microsoft provide details on internal company communications and research regarding the use of online content to train large language models. The filings, submitted September 4 in a case brought by Mother Jones and other outlets, quote statements from company executives and researchers on topics including data sourcing, output behavior, and potential effects on content creators. The cases remain in early stages, with defendants maintaining that their training methods qualify as fair use under U.S. copyright law.

The publishers allege that the companies incorporated copyrighted articles into training datasets without permission or compensation. Court records show that Microsoft director of applied science Brent Hecht wrote in an internal document that large models were ingesting work in ways that “millions of people around the world will soon consider” an “astonishing theft of unprecedented proportions.” The same document noted that “almost no one intended for content they created to be used in this fashion, nor are they compensated for its use.” OpenAI and Microsoft have not publicly confirmed or denied the specific phrasing but have stated in court papers that training on publicly accessible material constitutes transformative fair use.

Microsoft research cited in the filings reported that users of AI-assisted search tools clicked through to news articles at rates approximately 90 percent lower than users of conventional search engines. A Microsoft attorney described chatbots as systems “designed to answer your question” rather than directing users to external links. Separate analysis referenced in the documents indicated that referral traffic from Google Search and Google Discover to publishers declined by at least 30 percent after the introduction of AI-generated overviews. Company representatives have countered that any reduction in referrals reflects broader shifts in user behavior and that AI tools can increase overall engagement with information sources over time.

Hecht also warned in internal materials that reliance on AI-generated content risked creating a feedback loop that could “threaten the economic foundations of its essential suppliers.” The term “enshittification,” coined by writer Cory Doctorow to describe progressive degradation of online platforms, appears in the plaintiff brief as a characterization of this process. OpenAI president Greg Brockman, who has contributed to political committees supporting Donald Trump, exchanged messages with colleagues about the model’s ability to complete sentences from New York Times articles. In one exchange, Brockman responded “ah nice” to a suggestion about bypassing the Times paywall for scraping purposes. OpenAI has stated that such internal discussions reflect exploratory technical work rather than finalized policy.

Training datasets described in the filings included targeted collection of news articles, which formed a substantial portion of one version of the data. OpenAI expressed an objective to “crush freshness” in real-time news domains. The company also incorporated existing research datasets that contained paywalled articles, despite those datasets’ terms prohibiting commercial use. Microsoft chief executive Satya Nadella testified that companies should avoid accessing data in violation of terms of service. Researcher Glen Weyl, who established a Microsoft research unit, wrote that compensating creators aligned with the interests of the company and other stakeholders. The filings note that copyright notices, author names, and terms-of-use statements were removed during dataset preparation. A Microsoft document stated that “almost none of this content has been obtained with its owners’ knowledge or consent.”

After the lawsuits were filed, OpenAI implemented a filter that prevented its models from reproducing text from the suing publishers. The plaintiff brief states that no equivalent filter was applied to non-suing entities and that the measure’s purpose was to limit litigation evidence rather than to address copyright generally. A Columbia Journalism Review experiment found that ChatGPT declined to cite articles from the plaintiffs. Hecht cautioned that such filters could reduce visibility for rights holders into training data usage.

OpenAI analyzed 20 million ChatGPT responses for sequences of 25 consecutive words matching published articles. The review identified more than 400,000 such overlaps with Mother Jones content, a volume comparable to that found for the New York Times. The company noted that some overlaps may result from quoted material rather than direct reproduction of original reporting. Plaintiffs argue these findings demonstrate verbatim copying; defendants maintain that isolated n-gram matches do not establish infringement when the overall use is transformative.

The filings describe a collective-action problem in which individual firms benefit from using content without payment while the broader ecosystem of original reporting faces revenue pressure. The brief states that “while the industry as a whole would benefit if every company paid… each individual company is better off taking content for free while others pay.” OpenAI and Microsoft have responded in court that licensing every piece of training data at scale would be impractical and that existing legal precedents support fair-use treatment of non-expressive uses such as model training. Similar arguments appear in parallel cases brought by The New York Times and The Intercept.

The litigation continues in federal court, where judges will evaluate whether the companies’ data practices meet the statutory factors for fair use. No final rulings on liability have been issued. Both sides have submitted additional expert declarations addressing the technical nature of model training and the economic impact on publishers.