The legal battle over how artificial intelligence companies use copyrighted material to train their systems is entering a critical stage, with newly unsealed court documents offering a detailed look at concerns raised inside Microsoft and OpenAI about their use of publishers’ work.
At the heart of the dispute is a question that has shaped the legal history of major technologies, from VCR recordings and book digitization to the use of existing computer code to develop new operating systems: whether the use of copyrighted material without permission is protected by fair use.
That same question has now become one of the central issues in lawsuits brought by publishers, studios and recording companies against AI companies. As large language models increasingly become tools through which billions of people may access information online, the outcome of these cases could have significant consequences for both the technology industry and the businesses that create copyrighted content.
The latest documents were unsealed Thursday as part of litigation involving Microsoft, OpenAI and The New York Times, along with 11 other publishers. They reveal internal discussions that appear to show substantial concern within the companies about the copyright and intellectual property implications of developing AI systems using large quantities of online content.
One internal Microsoft document from 2023 warned that people around the world could eventually view the large-scale collection of copyrighted material for AI training as “an astonishing theft of unprecedented proportions.” The document also acknowledged that most creators had not intended for their work to be used in this way and were not being compensated for it.
Microsoft Director of Applied Science Brent Hecht was particularly critical of the practice, describing the ingestion of copyrighted works as potentially “the largest theft of labor in human history,” according to the newly unsealed filings.
The documents could prove significant as the companies defend themselves against claims that their AI systems were developed through the unauthorized use of publishers’ copyrighted work.
Fair Use At The Center Of The Dispute
The legal fight ultimately turns on the doctrine of fair use, which permits certain uses of copyrighted works without obtaining permission from the copyright holder.
Courts generally examine four factors when determining whether a particular use qualifies as fair use. Among the most important are the purpose and character of the use, including whether it is commercial and whether the copyrighted material has been transformed into something serving a new purpose, as well as the effect of the use on the potential market for the original work.In the current litigation, Microsoft and OpenAI bear the burden of establishing that their use of copyrighted material falls within those protections.
The publishers, however, argue that the companies’ activities go beyond a transformative use of copyrighted material and instead involve copying and reproducing journalism in ways that compete directly with the original publishers.The issue of how the companies obtained access to certain content could also become important.
OpenAI Paywall Exchange Draws Attention
One of the newly highlighted exchanges involves OpenAI co-founder and then-president Greg Brockman.According to the court filing, an OpenAI engineer informed Brockman about a technical method described as a “hack to get around [the] nytimes paywall” while the company was exploring ways to scrape content from The New York Times website.Brockman’s response was, “Ah nice,” according to the filing.
The publishers argue that the exchange demonstrates awareness within OpenAI of methods for circumventing access restrictions placed on copyrighted material.They further allege that this was part of broader efforts by employees to develop ways of getting around paywalls without being detected.
The significance of the exchange is tied to Supreme Court precedent concerning fair use. Publishers are expected to argue that evidence of an “evasive motive,” particularly when accompanied by violations of industry standards or restrictions, can weaken a company’s fair-use defense.Microsoft’s position, however, appears more complicated than OpenAI’s in this area.
Microsoft CEO Satya Nadella testified that content protected by a paywall should be licensed by anyone seeking to use it, according to the filing. Nadella also said that if he had known OpenAI had scraped and used restricted material to train its models, he would have exercised Microsoft’s contractual rights to require the company to retrain those models.
The testimony could become an important point of comparison as the court evaluates the companies’ knowledge of how copyrighted material was obtained and used.
Publishers Argue AI Systems Compete With Their Journalism
Another major issue is the fourth factor of fair use: the effect that the alleged copying has on the existing or potential market for the copyrighted work.
The publishers contend that AI chatbots are not simply using news articles as background material to generate new information. They argue that the systems can reproduce substantial portions of original journalism, potentially allowing users to obtain information without visiting the publisher’s website.
The New York Times has specifically pointed to instances in which AI systems allegedly reproduced articles almost word for word when prompted.
In one example cited in the litigation, Bing Chat reproduced all but two of the first 396 words of a 2023 New York Times article titled “The Secrets Hamas Knew About Israel’s Military.”
Another exhibit reportedly identifies 100 instances in which OpenAI’s GPT system had trained on and memorized Times articles and subsequently produced portions of them word for word.
The publishers argue that such responses are fundamentally different from the limited snippets traditionally displayed by conventional search engines. They say AI-generated responses can provide enough of an article’s content to satisfy a user’s information needs without requiring the user to click through to the original publication.
That distinction could be particularly important to the market-harm portion of the fair-use analysis.
Internal Data Shows Decline In Search Clicks
The publishers are also pointing to evidence that AI-powered search products may already be affecting traffic to traditional news websites.According to internal Microsoft data cited in the filing, The New York Times experienced significantly lower click-through rates when users searched through Bing Chat compared with conventional web search.
The decline was reported to be between 87 percent and 93 percent.For publishers whose businesses depend heavily on readers visiting their websites, such a reduction could represent a substantial change in the economics of online journalism.
The publishers argue that the issue is not limited to traffic generated by existing AI products. They also contend that Microsoft’s and OpenAI’s use of copyrighted material without payment could interfere with emerging licensing markets in which publishers negotiate agreements allowing their content to be used by AI companies.
Licensing Deals Complicate The Companies’ Defense
The existence of licensing agreements between AI companies and other publishers is another issue likely to receive significant attention.Publishers suing Microsoft and OpenAI argue that the companies cannot simultaneously claim that there is no viable market for licensing copyrighted news content while entering into agreements with publishers willing to provide that material under negotiated terms.
The argument is particularly relevant to the fair-use analysis because the availability of a licensing market can be considered when assessing the economic impact of unauthorized use.
The publishers contend that if AI companies can obtain content through licensing agreements, their decision to use other copyrighted works without compensation can undermine the potential market for those licenses.That issue could also influence the broader debate over how AI companies should obtain training data as the industry moves toward increasingly sophisticated models.
OpenAI’s Own Statements Could Become Significant
The court filing also cites comments from Nick Turley, OpenAI’s head of ChatGPT, concerning the relationship between AI products and original content.According to the filing, Turley said, “Our products are largely substitutive, period.”The statement is notable because one of the key questions in the litigation is whether AI-generated responses complement original journalism or instead replace the need for users to access it.
If AI systems provide users with substantially the same information contained in an article, publishers argue that the systems can effectively compete with the original source. OpenAI and Microsoft, meanwhile, maintain defenses based on the legal framework governing the use of copyrighted material.The competing arguments place the fair-use doctrine under pressure as courts consider technology that can ingest enormous quantities of copyrighted works and generate new material in response to individual user requests.
A Broader Test For The AI Industry
The dispute between The New York Times, other publishers, Microsoft and OpenAI is part of a much larger legal debate over the use of copyrighted material to develop generative AI.The technology industry’s position is that training large language models involves a form of technological innovation in which information is processed to create systems capable of generating new responses. Copyright holders argue that the systems cannot be developed at their expense while reproducing their work and competing with the markets that support it.The newly unsealed documents add another dimension to that dispute by providing evidence of internal discussions about the legal and ethical implications of collecting copyrighted material.
The outcome of the litigation could help determine how courts apply fair-use principles to AI training and how much weight should be given to factors such as commercial purpose, transformation, access restrictions, reproduction of copyrighted material and damage to existing or potential licensing markets.The implications extend beyond news publishers. Similar questions are being raised by authors, visual artists, musicians and other creators whose work has been incorporated into AI training datasets.
As AI companies continue developing systems that increasingly serve as gateways to information, the legal boundaries surrounding their training data could ultimately determine how the next generation of internet technology interacts with copyrighted work.
For now, the central question remains whether the massive-scale use of copyrighted material to build AI systems represents a transformative technological use protected by fair use or an unauthorized exploitation of creative and journalistic labor. The newly unsealed evidence ensures that question will remain at the center of the legal battle.
