We introduce orphaned sophistication, a novel linguistic feature for distinguishing AI-generated prose from human-authored text. Orphaned sophistication occurs when a figurative expression (metaphor, personification, synecdoche) appears without the network of supporting choices — lexical, syntactic, tonal — that would normally license it in skilled human writing. We operationalise the feature through a three-component annotation scheme (Structural Integration, Tonal Licensing, Lexical Ecosystem) and present a hand-annotated corpus of 400 passages drawn equally from literary fiction and LLM outputs across four model families. Inter-annotator agreement is high (Cohen's kappa = 0.81). A logistic-regression classifier using only orphaned-sophistication scores achieves 78.2% balanced accuracy, and when combined with existing stylometric baselines the feature yields a statistically significant 4.3 percentage-point improvement (p < 0.01). Qualitative analysis confirms that the feature captures a failure mode — figurative language produced without underlying rhetorical commitment — that surface-level detectors miss. We release the annotation guidelines, corpus, and scoring code to support further work on interpretable AI-text detection.
Richard Quinn (Sun,) studied this question.