[Hot take] Problems with AI prose

A viral analysis on LessWrong by user 'sudo' has dissected a recent New York Times quiz that pitted AI-generated prose against famous human authors. The quiz presented five head-to-head comparisons where participants blindly chose their preferred excerpt. The results were startling: Anthropic's Claude Opus 4.5 model, tasked with rewriting famous passages in its own voice, was preferred by 65% of quiz-takers over Carl Sagan's science writing from 'The Demon-Haunted World.' In other categories, preferences were nearly split, with AI tying 50/50 on Cormac McCarthy's 'Blood Meridian' and losing only slightly in historical fiction (44% vs. 56% for Hilary Mantel).

The author, who strongly preferred all human writing, conducted a technical breakdown to highlight the deficiencies in AI prose. Using McCarthy's passage as a prime example, they argue the human text employs a powerful, intentional metaphor comparing war to stone—enduring, ever-present, and waiting for humanity. In contrast, Claude's output followed a simple, linear narrative without attempting such layered analogies. The core complaint is that while AI writing is competent and often preferred in blind tests, it lacks the deep, intentional craftsmanship, metaphor, and visceral imagery that define great literature. The post serves as a warning: if the quality of AI writing does not improve beyond surface-level fluency, the proliferation of such content risks degrading our broader literary culture.