OpenAI says the Times cheated to get ChatGPT to copy articles

If you spend hours engineering ultra-specific prompts just to force a chatbot to recite a paywalled article word-for-word, did the AI steal it, or did you just break it on purpose?

That is essentially OpenAI's defense. In a court filing on Monday, the AI startup asked the US District Court in Manhattan to dismiss key portions of The New York Times' high-stakes copyright lawsuit, arguing the newspaper "cheated" to manufacture evidence.

OpenAI calls NYT prompts a "hack"

According to OpenAI, The Times had to resort to deceptive prompts and exploit a known bug to get ChatGPT to regurgitate verbatim excerpts of its articles.

OpenAI argued that developers are already working to patch the flaw, pointing out that no regular user actually interacts with AI this way. To drive the point home, OpenAI even cited a Times article from April 2023, titled "35 Ways Real People Are Using A.I. Right Now," claiming real-world usage looks nothing like what the outlet presented in court.

The Times fired back immediately. Lead counsel Ian Crosby told The Verge that framing their investigation as a hack is absurd. In Crosby's words, the newspaper was simply using OpenAI's own tools to find proof that their copyrighted reporting was stolen and reproduced without consent. He also stressed that OpenAI has not denied copying Times works within the relevant statute of limitations.

Whittling down the lawsuit

OpenAI is asking the judge to throw out four of the seven claims in the lawsuit.

The targets for dismissal include contributory infringement, failure to remove copyright management info, and unfair competition claims. OpenAI is also trying to partially block the direct infringement claim by cutting off anything that occurred more than three years before the suit was filed.

It is a familiar playbook. OpenAI successfully gutted a copyright lawsuit from author Sarah Silverman and others using nearly identical arguments, whittling that case down to a single count of direct infringement.

However, taking down The Times will be a much steeper climb. The publisher has spent decades fighting intellectual property battles and has leveled serious allegations around trademark dilution and revenue destruction.

Why it matters

This legal sparring cuts right to the heart of generative AI's biggest existential question: does training an LLM count as fair use, or is it high-tech piracy?

For enterprise builders and developers, the outcome here will dictate the safety of the entire foundational model supply chain. If OpenAI fends off these claims, building on top of frontier models remains relatively insulated. But if courts decide that memorization equals direct infringement, the downstream costs of licensing data could reshape API pricing overnight.

We think OpenAI's "normal people do not use it this way" defense is clever PR, but a double-edged sword in court. By arguing that the chatbot only regurgitated full articles because of deceptive prompting, OpenAI effectively admitted that the training weights still harbor those copyrighted articles verbatim.

It looks like OpenAI is doing everything it can to shrink the battlefield down to a narrow debate on fair use. But with rivals like Anthropic and Stability AI also drowning in lawsuits, the industry can only patch outputs for so long before a judge decides whether the training process itself was illegal.