There's no good way to validate that a model aligns with human values. The closest thing we have now would be extensive behavioral tests; e.g. does the chatbot

There's no good way to validate that a model aligns with human values. The closest thing we have now would be extensive behavioral tests; e.g. does the chatbot say things that a majority of cultural members strongly agree with in the vast majority of cases, including in large numbers of nuanced situations where the context matters and differentiates a Western vs. non-Western response. No one knows how to make those kinds of behavioral tests comprehensive, either.