Late Monday, The New York Times responded to OpenAI’s claims that the newspaper “hacked” ChatGPT to “set up” a lawsuit against the leading AI company.
“OpenAI is wrong,” The Times repeatedly argued in a court filing opposing OpenAI’s motion to dismiss the NYT’s lawsuit accusing OpenAI and Microsoft of copyright infringement. “OpenAI’s attention-grabbing claim that The Times ‘hacked’ its products is as irrelevant as it is false.”
OpenAI had argued that NYT allegedly made “tens of thousands of attempts to generate” supposedly “highly anomalous results” showing that ChatGPT would produce excerpts of NYT articles. The NYT’s allegedly deceptive prompts—such as repeatedly asking ChatGPT, “what’s the next sentence?”—targeted “two uncommon and unintended phenomena” from both its developer tools and ChatGPT: training data regurgitation and model hallucination. OpenAI considers both “a bug” that the company says it intends to fix. OpenAI claimed no ordinary user would use ChatGPT this way.
But while defending tactics used to prompt ChatGPT to spout memorized training data—including more than 100 NYT articles—NYT pointed to ChatGPT users who have frequently used the tool to generate entire articles to bypass paywalls.
According to the filing, NYT today has no idea how many of its articles were used to train GPT-3 and OpenAI’s subsequent AI models, or which specific articles were used, because OpenAI has “not publicly disclosed the makeup of the datasets used to train” its AI models. Rather than setting up a lawsuit, NYT was prompting ChatGPT to discover evidence in attempts to track the full extent of copyright infringement of the tool, NYT argued.
To figure out if ChatGPT was infringing its copyrights on certain articles, NYT “elicited examples of memorization by prompting GPT-4 with the first few words or sentences of Times articles,” the court filing said.
