Researchers at the University of North Carolina at Chapel Hill have identified a persistent limitation in artificial intelligence creative writing: while AI generates increasingly fluent narratives, its characters consistently lack the ambiguity and complexity that define compelling human fiction. The findings, derived from a new automated evaluation framework, underscore that current language models prioritize narrative efficiency over emotional depth. The study introduces CASPER, an automated benchmarking system designed to analyze thousands of AI-generated stories across eight literary dimensions. These metrics assess character realism, development over time, reliance on archetypes, and the preservation of narrative mystery. By systematically applying literary theory to machine-generated text, CASPER provides the first standardized method for measuring character depth in artificial intelligence fiction. Analysis reveals that AI writing tools consistently play it safe, resolving character arcs with tidy conclusions and leaning on familiar tropes. Human authors, by contrast, frequently allow characters to remain contradictory or unresolved, an approach that literary research shows often leaves a stronger psychological impact on readers. Lead author Anneliese Brei noted that this willingness to embrace ambiguity is a key differentiator between machine and human storytelling. The research also challenges the assumption that model scale automatically translates to creative quality. Co-author Nicholas Sanaie found that larger, more computationally intensive models do not produce significantly more diverse or nuanced characters than smaller variants, indicating that the bottleneck lies in narrative understanding rather than processing power. The study arrives as AI writing assistants like Sudowrite and Squibler gain adoption among novelists, and AI-generated scripts increasingly enter film and television development pipelines. Industry surveys indicate a growing subset of writers now incorporate machine learning into their drafting processes, making independent evaluation tools essential. Senior author Snigdha Chaturvedi emphasized that CASPER offers developers a concrete metric to track whether newer models are genuinely improving in character construction or merely expanding vocabulary and syntactic fluency. For creative professionals, the research outlines a clear division of labor. Artificial intelligence remains a highly capable drafting partner, but it struggles to replicate the deliberate uncertainty that makes fictional personas resonate. Developers aiming to build next-generation storytelling tools must prioritize narrative complexity and emotional ambiguity alongside generative fluency. As AI continues to integrate into entertainment and publishing workflows, frameworks like CASPER will likely serve as industry standards for assessing whether machine-generated narratives can match the psychological realism of human-authored fiction.