Seems fair. Is that really how the training works though? You'd hope they'd have a first-level filtering process that flags things before they get fed into the model but who knows.