That's really the core of the issue - if your prompt/input contains 10 bytes of salient information, and the output is a 1kB document, all you've done is add 99

That's really the core of the issue - if your prompt/input contains 10 bytes of salient information, and the output is a 1kB document, all you've done is add 990 bytes of fluff. Commonly referred to as "AI slop". If the goal is boilerplate code, or a fuzzy view of documentation from the corpus, or "pretty" images to gawp at in private, that's fine and dandy, because you're just extracting and viewing the model data on your own time. If you're "proofreading" an already-1kB letter for style / formality that's fine, since your input intents are high-fidelity and you're essentially referencing a style guide for some minor edits. But low-input generations are entirely inappropriate for crafting a message to other humans, because the result is 99% shoving the model in their face. Whether it's letters, pull requests, graphical art, or music - it's all communication, and low-input AI generations are just spam in that context, semantically equivalent to sending them the prompt instead of the output. And people can tell, because we're excellent pattern-matchers, and every bit of information/intent abdicated to the model will inevitably leave its mark on the output.