I think the question is, how much risk is involved and how much do those mitigating methods reduce it? And with that, we can figure out what applications it is appropriate for.

There are a lot of services out there that offer these types of AI guardrails, and it doesn’t have to be expensive.

Not saying that this approach is foolproof, but it’s better than relying solely on better prompting or human review.