Start by figuring out if the systems organizations build around AI are designed to produce trustworthy outcomes. That's an architectural question, not a model question.