The causal mask ensures each token can only attend to itself and previous tokens, preventing a language model from peeking at future positions.