Background & Context§
The New York Times v. OpenAI and Microsoft copyright lawsuit, filed three years ago, has become a bellwether for how generative AI companies source training data. The case centers on allegations that the defendants used millions of Times articles without authorization to train large language models (LLMs) like GPT-4 and Copilot. On September 17, 2026, new unredacted filings surfaced, containing explosive internal communications. According to The Times' brief, a top Microsoft executive described AI scraping as 'theft,' and OpenAI leadership acknowledged an 'existential threat' to publishers. This matters because it undercuts the industry's longstanding fair use defense, suggesting even insiders knew their practices were legally and ethically dubious.
The News: What Happened Exactly§
The unsealed material, part of a three-year-old lawsuit, reveals that a senior Microsoft executive privately characterized the companies' AI training practices as 'theft'—specifically, 'the largest theft of labor in human history.' This admission, buried in unredacted filings, directly contradicts public statements by AI firms that scraping publicly available web data is fair use. The Times' brief further alleges that OpenAI's own leadership internally described its models as an 'existential threat' to the publishers and journalists whose work trained them. These quotes come from The Times' legal brief, not the underlying exhibits, which remain sealed—meaning the full context is unavailable and could alter interpretation. Nonetheless, the filings paint a picture of corporate awareness that their data acquisition methods were legally precarious.
The unredacted documents also detail the mechanisms allegedly used to obtain content. According to the filing, the companies bypassed paywalls undetected, built training datasets via mass scraping, and deliberately stripped copyright notices from training data. Bypassing paywalls typically involves spoofing user agents or using credentialed access without permission—techniques that, if proven, could constitute circumvention of technological protection measures under the Digital Millennium Copyright Act (DMCA). Stripping copyright notices, often done via automated preprocessing scripts, removes metadata that would otherwise flag provenance. A simplified example of such a pipeline might look like:
import requests
from bs4 import BeautifulSoup
def scrape_article(url, headers={'User-Agent': 'Mozilla/5.0'}):
response = requests.get(url, headers=headers)
soup = BeautifulSoup(response.text, 'html.parser')
# Remove copyright notices and footer metadata
for element in soup.find_all(class_=['copyright', 'footer', 'legal']):
element.decompose()
return soup.get_text()This pseudocode illustrates how copyright management information (CMI) could be systematically removed, a practice that, if intentional, violates DMCA Section 1202. The filing does not specify the scale, but previous estimates suggest OpenAI trained GPT-3 on ~300 billion tokens, with a significant portion from news domains. The legal implication is stark: if the companies knowingly stripped CMI, they could face statutory damages of up to $25,000 per violation under 17 U.S.C. § 1203. Moreover, the 'theft' admission could be used to argue willfulness, potentially tripling damages.
Historical Parallels & Similar Incidents§
This is not the first time a technology giant has been accused of exploiting content creators. In 2004, Google launched its Book Search project, scanning millions of library books without permission. Authors and publishers sued (Authors Guild v. Google), and in 2015, the Second Circuit ruled that Google's snippet view was transformative fair use. However, the key difference is that Google made digitized copies available for search, not for training generative models that could reproduce expression. In the current case, the alleged stripping of copyright notices and paywall bypassing goes beyond indexing—it directly enables model training that competes with the original content. The lesson from Google Books is that fair use hinges on transformativeness; but when a model can generate substitute articles, the transformation argument weakens.
Another parallel is the 2013–2015 period when Facebook manipulated user feeds for emotional contagion experiments, and later, the Cambridge Analytica scandal. Those incidents showed that internal memos and private communications can become devastating evidence of intent. Similarly, the Microsoft executive's 'theft' remark could be the smoking gun that shifts the legal narrative from 'fair use' to 'willful infringement.' In both cases, the technology companies initially denied wrongdoing, but internal awareness of ethical breaches undermined their public defense. The contrast here is that AI scraping directly threatens the economic viability of journalism—unlike social media experiments, which affected user sentiment but not livelihood. As the Times argues, if AI models are trained on paywalled journalism and then used to generate competing summaries, the original publishers lose both traffic and licensing revenue.
A more recent parallel is the 2023 lawsuit by Getty Images against Stability AI, which alleged that the company scraped millions of copyrighted images to train Stable Diffusion. That case, still ongoing, similarly relies on internal documents showing that Stability AI's founders acknowledged the legal risks. The lesson from all these incidents is that courts increasingly scrutinize the process of data acquisition, not just the output. If the Times can prove that OpenAI and Microsoft deliberately bypassed paywalls and removed copyright notices, the fair use defense may collapse—not because training is inherently infringing, but because the means were illicit. For developers, this underscores the importance of data provenance and compliance; for founders, it signals that internal communications can become Exhibit A.
The Broader AI Landscape§
The unredacted filings arrive amid a global regulatory push. The EU AI Act now requires disclosure of training data sources, and the U.S. Copyright Office has issued guidance that AI-generated outputs may not be copyrightable. If the Times prevails, it could set a precedent requiring licenses for training data, similar to how music streaming services pay royalties. This would fundamentally alter the economics of LLM development, where data is currently treated as free raw material. Startups relying on scraped data may face existential cost increases, while incumbents with licensed archives (e.g., Bloomberg, Reuters) could gain an advantage. The 'theft' quote also fuels public sentiment that AI companies are extracting value from creators without compensation—a narrative that could accelerate legislative action.
What to Watch§
- The underlying exhibits: Still sealed, but if unsealed, they may provide context that mitigates or amplifies the quotes.
- Judge's ruling on fair use: A decision could come in 2027; appeals are likely.
- Industry response: Expect AI companies to tighten data provenance and strike licensing deals proactively.
- Regulatory fallout: The FTC and state attorneys general may open investigations into unfair competition.
The case is a defining moment for generative AI's legal foundation. As the filings show, even insiders harbored doubts. For developers and founders, the message is clear: data sourcing is no longer a back-office concern—it is a board-level risk.