As I said, we are continuing to research the science and engineering of how to increase accuracy and reliability. In a funny way, it's not directly using the content to return that content to the user. It's using it to determine the weights of billions, in some cases trillions, of parameters that allow it to make more accurate estimates as to what the next words in a response should look like.
On April 20th, 2023. See this statement in context.