Perceiving the humanness: reader evaluation of AI-generated versus human-written romantic texts

The rapid advancement of generative AI (GenAI), particularly large language models (LLMs) like GPT-4o, has revolutionized content creation, enabling efficient eBook production. However, this raises concerns about authorship ambiguity, quality, and the proliferation of low-quality AI-generated content on Amazon Kindle Direct Publishing. Prior research on AI-text detection has focused on logical, academic, or short texts (e.g., AI-generated messages and hotel reviews), leaving a gap in understanding AI's capability to mimic human writing in emotionally rich genres like romance, which relies on nuanced emotional expression, authenticity, and tone of voice. While AI can generate content quickly in massive amounts with high-cost efficiency, its ability to replicate the emotional depth and authenticity of human-written romance narratives remains uncertain. This creates challenges for readers, platforms, and regulators in assessing authorship and ensuring content quality. Which leads to the research question: What are readers' perceptions of AI-generated and human-written romantic narratives based on (a)linguistic naturalness, (b)coherence, and (c) emotional tone? A mixed-methods approach was employed, combining quantitative surveys (5-point Likert scales) and qualitative thematic analysis. Participants (n = 85) evaluated two 170-word romantic reunion passages: Text A (Chapter 2 from Nicholas Sparks' The Notebook) and Text B (GPT-4o-generated). Metrics included linguistic naturalness (vocabulary simplicity, sentence structure), coherence (logical flow, consistency), and emotional tone (authenticity, intensity). Text order was randomized to minimize bias. More specifically, quantitative analysis involves paired-sample t-tests and MANOVA tests, coupled with Shapiro-Wilk and Wilcoxon tests analyze survey responses. And qualitatively, thematic analysis of open-ended explanations is used. The results showed that only 51.8% correctly identified the human text (Text A), with a marginal advantage (11.8%) over AI text attribution, indicating difficulty in distinction. In terms of linguistic naturalness, AI text used significantly more diverse vocabulary ( d = 1.32, p < .001), longer sentences (d = 1.58, p = .002), and more descriptive details (d = -0.35, p = .033). Human text was qualitatively associated with simpler vocabulary and sentences. With coherence metrics, no significant differences emerged in ease of reading, logical consistency, or transitions (p > .05), which reinforces the importance of developing genre-specific frameworks. The measurement of emotional tone was quantitatively indistinguishable, but qualitative analysis revealed that AI text showed more emotional intensity, while human text was more authentic. Lastly, em-dashes emerged as a prominent indicator of AI-writing by participants, serving as a novel detection cue. In conclusion, genre is a critical aspect in AI-text detection: romance narratives demand metrics beyond lexical or coherence markers, but more emotional and micro-punctuation measures. GPT-4o narrows the human-AI gap in romantic narration but struggles to replicate emotional authenticity. Em-dashes emerged as a reliable, non-theoretical AI indicator, showing systematic biases in training data. A mixed-method approach is essential as both measures provide a comprehensive insight into this study. Practical implications demand that writers prioritize authentic storytelling, educators teach critical reading, and developers refine LLMs by suppressing robotic markers.