Making sure the model stays aligned over time, training, and model revisions is a separate issue.
Asking the model gives you a snapshot of its current perception of words which can vary over time. Let a model cosplay as a super villain and it will treat "good" and "evil" as something different
