At worst, it's going to lead to a "is this a good NPC response" pipeline, where the model spits out lines based on its given prompt, and each one is approved by

At worst, it's going to lead to a "is this a good NPC response" pipeline, where the model spits out lines based on its given prompt, and each one is approved by a proofreader. The minutes of time spending on thinking up what an NPC would say is now (about 10 seconds of speedreading 5 lines of generated results + 5 seconds in clicking either the 'approve' or 'reject' button): That in of itself is at least a 10x increase in available topic queries for NPCs.