LLMs have detectable stylistic fingerprints that don't disappear with prompt engineering or decoding tweaks—the model itself and genre are far more important than generation settings in shaping text style.
This paper analyzes stylistic differences between human-written and LLM-generated text across 11 models, 8 genres, and multiple decoding strategies using linguistic features. The key finding: LLM writing has consistent stylistic markers that persist regardless of prompting tricks or decoding settings, and genre matters more than whether text is human or machine-written.