The post where I found this was a continuation of a thread that started with a post[0] by Terence Tao saying:
"In recent weeks there have been a number of examples of Erdos problems that were solved more or less autonomously by an AI tool, only to find out that the problem had already been solved years ago in the literature"