I'm creating #syntheticdata for teaching in the social sciences & find that #SDG with LLMs isn't for my small-scale use. While there are workflows to combine LLMs & generate more credible output ( https://link.springer.com/chapter/10.1007/978-3-031-93418-6_9 ), general-purpose models often create results that are too diverse & reflexive, even when imitating oral communication. Such data reminds me of journalism scandals à la Stephen Glass. High-quality data in my case is more messy and dull. Just look at YouTube comment sections.