OpenAI and The New York Times are now fighting not only over copyright, but over how the evidence in the case was produced. In a court filing Monday, OpenAI argued that the paper’s examples of ChatGPT reproducing Times material came from abnormal testing, not ordinary use of its products.
The Times rejects that account. Its lawyer told Ars that the testing was simply an effort to find whether OpenAI’s products could reproduce copyrighted Times work.
OpenAI’s Core Argument
OpenAI claims The New York Times paid someone to hack products like ChatGPT in order to set up its lawsuit. The company says the paper presented about 100 examples in which some version of OpenAI’s GPT-4 model generated several paragraphs of Times content after user prompts.
According to OpenAI, those examples do not show how normal people use ChatGPT. The company says The Times needed tens of thousands of attempts to produce what it called highly anomalous results by targeting and exploiting a bug that OpenAI says it is committed to addressing.
OpenAI describes the testing as contrived attacks by a hired gun. In its telling, the effort pushed models until they either hallucinated fake NYT content or reproduced training data in ways that resembled Times articles.
That argument matters because OpenAI is trying to narrow the lawsuit. The company asked the court to dismiss most of The Times’ claims, including allegations tied to direct copyright infringement, contributory infringement, Digital Millennium Copyright Act violations, and misappropriation.
Why The Times Disputes The Hacking Claim
Ian Crosby, a Susman Godfrey partner and lead counsel for The New York Times, told Ars that OpenAI was mischaracterizing the paper’s conduct. He said what OpenAI called hacking was simply using OpenAI’s products to look for evidence that they stole and reproduced The Times’s copyrighted works.
Crosby also said OpenAI’s filing does not dispute that the company copied millions of The Times’ works to build and power its commercial products without permission. The Times’ position is that the problem is broader than the 100-plus examples included in its complaint.
The Times also pushes back against OpenAI’s timing argument. OpenAI said The Times had reported on chatbot developments for years and did not raise copyright concerns after OpenAI disclosed in 2020 that Times articles were used to train its AI models. OpenAI says The Times only acted after ChatGPT’s rapid adoption following its debut in 2022.
OpenAI’s filing says The Times reached out to demand commercial terms after reports about the value unlocked by the technology, then filed suit two days after Christmas and sought billions of dollars. Crosby told Ars that OpenAI is now saying it is too late to bring claims even though, in The Times’ view, OpenAI was secretive and deliberately concealed how its products operate.
What The ChatGPT Examples Involve
The dispute centers on two behaviors OpenAI says The Times deliberately targeted: training data regurgitation and model hallucination. OpenAI says both appeared in outputs from its developer tools and ChatGPT.
OpenAI described regurgitation as a situation where AI tools generate a sample that closely resembles training data. The company said this most often happens when the training dataset contains highly similar observations, such as duplicates of a particular work.
OpenAI says The Times used prompts that asked for the opening paragraph of a particular article, followed by repeated requests for the next sentence. But OpenAI argues that even this tactic did not generate entire articles. Instead, it says the outputs were scattered and out-of-order quotes.
The company also accused The Times of using ellipses in a way that obscured the order of the snippets and created the impression that ChatGPT had produced sequential, uninterrupted article text.
The hallucination examples are different. OpenAI says hallucinations happen when a model produces seemingly realistic answers that turn out to be wrong. In this case, OpenAI says The Times cited examples where AI models invented articles that The Times never published.
OpenAI argues that because none of the generated links to those bogus articles worked, any user receiving that output would immediately recognize it as a hallucination.
The Legal Stakes
If OpenAI’s motion is granted, the case would be substantially narrowed. OpenAI says some claims fail because they are time-barred, especially claims seeking damages tied to training data for older models. It also argues that other claims misunderstand fair use or are preempted by federal laws.
The Times’ response is that OpenAI is trying to avoid accountability for copying within the statute of limitations to train more recent and current models. Crosby told Ars that building new products is not an excuse for violating copyright law.
The case also raises a product question. OpenAI says ChatGPT is not a substitute for a New York Times subscription and that ordinary users cannot use ChatGPT to serve up Times articles at will. The Times says the examples show a deeper issue: that OpenAI’s systems can reproduce copyrighted works and generate misinformation.
OpenAI says reducing and eventually eliminating hallucination is an ongoing challenge in AI development. The company ties that effort to the use of more complete training datasets to improve model predictions.
Crosby told Ars that OpenAI’s filing also appeared to show that the company is tracking users’ queries and outputs, which he said was surprising because OpenAI had claimed not to do so. He said The Times expects to explore that issue in discovery.
What Comes Next
The immediate question is whether the court accepts OpenAI’s attempt to dismiss most of the claims. If it does, The Times’ case would move forward in a narrower form. If it does not, the broader fight over ChatGPT, copyrighted training data, regurgitation, hallucination, and fair use continues.
For now, the two sides are telling sharply different stories about the same evidence. OpenAI says the examples were engineered through abnormal prompting and exploitation of a bug. The Times says the examples reveal what its lawsuit is about: whether OpenAI copied and reproduced its copyrighted journalism without permission.