A distribution shift happens if the data the machine learning model is trained on has a different distribution than the data the model will be given in production.