"What was new was that the stories had been taken from legitimate news outlets and modified using large language models, most likely one built by Openai" Huh?

"What was new was that the stories had been taken from legitimate news outlets and modified using large language models, most likely one built by Openai" Huh? What is the evidence for this claim? I assumed they used an open weights model to avoid a distracting adversarial cat and mouse game against OpenAI.