OpenAI is pushing back against The New York Times in their ongoing copyright dispute by focusing on how the newspaper produced examples of article-like output from GPT models.
In a legal filing, OpenAI alleges that an individual paid by the Times used deceptive prompting to generate copies of New York Times articles. The Times’ position, according to its lawyer, is that the work was not hacking but a search for evidence of copyrighted content inside OpenAI’s AI models.
What OpenAI is alleging
OpenAI’s argument centers on the way the outputs were created. The company says the Times relied on prompts that violated OpenAI’s terms of service and were designed to make the models produce verbatim passages.
According to OpenAI, the results cited by the Times were not ordinary responses from its products. The filing describes them as highly unusual outputs produced only after extensive attempts.
OpenAI claims the Times paid someone to hack its products through prompting. It also says the work involved targeting a bug, using prompts it calls deceptive, and feeding the tool portions of the same articles that the prompts were trying to reproduce.
The filing says the process took tens of thousands of attempts to generate the results that became Exhibit J to the complaint. OpenAI also claims that virtually all of the passages involved already appear on multiple public websites.
Why the prompts matter
The dispute is not only about whether GPT models can produce copies of New York Times content. It is also about how those copies were produced and what that means for the copyright claims.
The Times demonstrated in its lawsuit that OpenAI’s GPT models could generate copies of NYT articles. OpenAI is now trying to frame those examples as the product of unusual and improper prompting rather than normal model behavior.
That distinction matters because OpenAI is describing the results as provoked. The company has argued that the alleged regurgitation was a rare error in the systems’ learning process and that the bug could be fixed.
OpenAI had already accused The New York Times earlier this year of using manipulative prompts to deliberately provoke copyright infringement by its AI models. The new legal filing develops that same argument in more direct terms.
The Times’ response
The New York Times’ lawyer rejects OpenAI’s description of the prompting as hacking. The lawyer claims the work was simply an effort to look for evidence of copyrighted content in OpenAI’s AI models.
That response puts the same activity in a very different light. OpenAI describes deceptive prompting and terms-of-service violations. The Times’ side describes evidence gathering for a copyright case.
The disagreement leaves several core questions at the center of the dispute:
- Were the outputs ordinary evidence of model behavior, or highly anomalous results?
- Did the prompts expose a copyright problem, or exploit a bug?
- Should feeding article portions into a model affect how the resulting passages are interpreted?
- Does a terms-of-service violation change the value of the evidence?
The source article does not resolve those questions. It presents OpenAI’s allegations and the Times lawyer’s counterargument as part of an ongoing legal fight.
What this says about AI copyright disputes
The case shows how much weight can fall on the method used to test an AI model. In a copyright dispute, the output itself is only part of the story. The prompt, the number of attempts, and the surrounding process can become central facts.
OpenAI’s position is that the Times did not merely ask its models simple questions and receive copied articles. The company says the newspaper’s examples depended on repeated attempts, targeted prompts, and article fragments supplied to the tool.
The Times’ side, as described in the source, argues that this was evidence collection. From that view, prompting a model to reveal copyrighted material is part of investigating whether such material exists in the system.
For readers following AI copyright litigation, the key takeaway is that the arguments are now about both content and conduct. The dispute covers what GPT models generated, but also how those generations were obtained and whether that process should count against the evidence.
OpenAI says the truth will come out in the course of the case. Until then, the public record described here leaves two sharply opposed narratives: one about deceptive prompting that exploited a bug, and another about searching for evidence of copyrighted material in AI systems.