You have to remember that LLM's don't have any persistent capacity to hold a "judgement". You ask for something, it provides an attempt at a completion for it. No fact checking, no reasoning, just a plausible looking output, tuned to hopefully get you to repeat the interaction.
Half the reason the dominant UX is a "Chat" is that's the only way to provide a facsimile of memory or persistence across requests. Append the last few turns, press go. Over time you can develop an eye for the model's tics/attractor topics.
Remember that they bill by token use, and suddenly, the entire UX/architecture starts making sense.
