In the decoder, attention scores for a word like 'am' have values for itself and all other words before it, but zero attention scores for future words like 'fine', which tells the model to put no focus on those words.