Self-attention as described so far lets every position see every other position. For a model that generates text left to right, that is a problem: predicting the next token while already looking at it teaches nothing. The causal mask enforces the rule that a token may look at the past and itself, never the future.
The mechanism is a single addition. Before softmax, the score from query i to key j is left alone if j is at or before i, and set to negative infinity otherwise. The formula on the card states it exactly: mask_ij equals 0 if j is less than or equal to i, else negative infinity. Softmax turns negative infinity into a weight of zero, so the masked positions contribute nothing. The grid shows the result, a lower-triangular pattern where each row i can attend only up to column i, with the diagonal marking the causal boundary.
Why this earns its own card is the training payoff in the second box. Without the mask you would have to feed the model one prefix at a time to avoid leaking the next word, which is slow. With it, the model can process the entire sequence in a single pass and still be trained to predict every position, because each position only ever saw legitimate earlier context. The card lists the three consequences: it prevents leaking future tokens during training, the same masked model behaves identically at inference where the future genuinely does not exist yet, and it enables parallel training over all positions.
This is the mask that makes GPT-style decoders trainable at scale. It is worth separating from the next card, the padding mask, which solves a different problem: that one hides filler tokens added to make batched sequences the same length, while this one hides the future.