When generating an image. The one you download they take it as positive feedback . My theory is that if you repeat a prompt twice or more they probably can label it as bad result. They could also use the enlarging of pictures after they are generated as additional feedback