@michaelgemar @cabel that's the problem with AIs: by definition, they can't be tested entirely, because nobody really knows what they will reply to a given request. They only give replies based on some probabilities ("predictions") that have been computed from data before.
@Deuchnord @cabel Right, but these issues crop up *so* quickly after their release — are they not testing at all, or at least some obvious edge cases? I feel like no one is internally “red teaming” these models.