If that's the case (only 0.1%), the developers must have done something that differs from other openai experiments that suggest code sequences that I recall seeing, where significant chunks of code from Stack Overflow or similar sites were appearing in answers.