Study Finds AI-Generated Novels Show Compressed Formal Variation

A recent study published on arXiv (2608.12630) examines if large language models can achieve the same level of formal diversity as that seen in human-generated texts. The research compares six different corpora: twenty novels crafted with GPT-5.5 in a nineteenth-century British realist style, twenty novels created with Qwen3-14B in the same style, another twenty novels from each model in a contemporary zero style, 205 human-authored British novels from the nineteenth century, and sixty-five contemporary human-written Zero-Style novels. The analysis evaluates metrics like MATTR-500, Shannon entropy, average sentence length, readability, and punctuation rate. Results show that AI-generated works display less formal variation than human texts, indicating that while models can imitate styles, they lack diversity in their outputs. This study adds to the ongoing dialogue regarding AI's capabilities and limitations in creative writing.