My personal guess is that they don't put it in training data, BUT they still have human read what you send, to see 1. what are the innovative uses that they can try to copy/integrate; 2. optimize (both in score and throughput) for their consumer's usage.