A generative model for antibodies is trained to produce sequences that look like the ones it was trained on. In practice it learns a probability distribution over antibody sequences — roughly, which amino acids tend to appear at which positions and in which combinations — and then samples new sequences from that distribution.
The sampling step is what makes the output useful. If the model only reproduced training sequences exactly, it would add nothing over a database search. Instead it produces variants: sequences that follow the same statistical patterns but differ in specific positions. Those differences are the design proposals. Because the model was trained on antibodies that fold and function, its samples tend to stay in the region of sequence space where antibodies are viable, which is exactly the region that random mutation would mostly miss.
The limitation is equally direct. The model has learned what antibodies look like, not what binds your particular target. Its samples are plausible antibodies, and plausibility is not the same as binding. This is why generation is only the first stage: a model that proposes a thousand reasonable sequences has not told you which one, if any, will work.