Where Models Concentrate and Humans Spread: Toward Cultural Reach in Generative AI

When a language model writes Harry Potter fanfiction, it reliably produces something recognizable: canon settings, familiar characters, conventional relationship patterns. What it rarely produces is the writing that fan communities distinctively produce---the stylistically irregular, fandom-internal, relationship-diverse work that makes up the periphery of the distribution but the heart of the culture. This gap is not a fluency problem. It is a coverage problem, and current evaluation frameworks have no way to measure it. We propose a human-grounded coverage framework that treats the empirical distribution of human responses as the standard for cultural AI evaluation---not whether outputs avoid failure, but whether they achieve genuine cultural reach. Our two metrics, LLM Coverage (LLM-Cov) and In-Boundary Rate (IBR), separate a model's cultural plausibility from its cultural breadth. Across ideation and narrative tasks, current LLMs are reliably plausible but systematically narrow: they concentrate near the cultural center of human response space and under-reach its periphery. Coverage makes this visible---and offers a positive standard for cultural AI: not whether a model can reproduce the mainstream, but whether it can reach the full range of meanings a cultural community produces.