The clearest differences between AI and human antibody design appear in three dimensions at once: how fast candidates can be produced, how many can be considered, and how far they stray from what is already known.
Speed and scale are linked. A human design cycle is limited by physical operations — immunizing an animal, growing cells, panning a library, expressing and assaying a variant. Each round of affinity maturation costs weeks and yields a small number of measured variants. An AI pipeline replaces the physical round with computation: a generative model can sample thousands of sequences and a scoring model can rank them in a fraction of that time. The asymmetry is not that AI is slightly faster; it is that AI searches a space that is too large to search physically.
Novelty is a separate axis and is easy to overstate. A generative model samples from a distribution learned from known antibody sequences, so its output is novel in the sense of being a new combination of learned patterns, not in the sense of being unrelated to anything in the training data. Novelty that falls far outside the learned distribution is exactly where the model's predictions become least trustworthy.
Reliability differs in kind. A human-derived candidate has already passed a real binding assay — the evidence and the candidate arrive together. An AI candidate arrives with a predicted score and no experimental evidence, so its reliability is inherited from how closely the new case resembles the model's training conditions.