Yes, but I don't see that being done, likely because safety is not what rakes in the $$$, so there aren't any best practices. If OpenAI cannot prevent people from taking their own lives, then we can be sure that the technology for safety is severely underdeveloped at the moment.

I think that one could train a model only on, say, children's books, and actually curate the inputs, which is the opposite approach of what the current models have been doing. If I were to make a database of knowledge I would want to teach my children, would I let it ingest porn, violence and adult fantasy stuff?