OpenAI has been in a war-room for days searching for a match in the data, and they just came out with this without providing proof.

My cynical opinion is that the traning corpus has some small amount of data generated by OpenAI, which is probably impossible to avoid at this point, and they are hanging on that thread for dear life.