I get that, and agree, but do we have confidence that passing text through an LLM's somewhat arbitrary persona-description capacities will preserve the diversity? It seems totally possible that the model would be blind to important latent variables — partly because people themselves may be.