Their end-user privacy policy says "We will not use your Inputs or Outputs to train our models" (followed by carve-outs, which seem reasonable).