Grammar Correction and Phrasing Refinement Lower Sentence-Level Perplexity and Entropy of Human Formal Writing

Abstract

Many AI-text detectors rely on perplexity, that is, on how predictable a text is to a language model. A natural question is whether a human can produce perplexity and entropy values similar to those of AI-generated text. We conduct a preliminary sentence-level comparison of five writing conditions: original humanwritten text, grammar-corrected human text, phrasing-refined human text, text from GLM 5.3 prompted to mimic the human author, and text generated by GLM 5.3 Flash without any human sample. All 13 matched sentences were scored with Qwen3. The mean perplexity decreased from 39.59 for the original human response to 37.24 after grammar correction and 21.26 after phrasing refinement, compared with 26.92 for the GLM 5.3 mimicking response and 37.52 for the GLM 5.3 Flash response. Mean entropy (in bits) was highest for the original and grammar-corrected human texts (2.98 and 3.05) and for the AI-generated texts it was 2.59 and 2.73, while the human-refined text had the lowest value (2.25). Sentence-to-sentence entropy changes were negatively autocorrelated for all human-originated texts and positively autocorrelated for both AI-generated texts, although none of the AR(1) coefficients was statistically significant. These results indicate that a polished human-written formal response can reach lower perplexity and entropy than AI-generated text, which may lead to false positives in perplexity-based detection.

Keywords

• AI-generated text detection
• perplexity
• Shannon entropy
• human writing
• false positives