Detecting the Human in the Machine: A Forensic Linguistic Approach to Distinguishing Human, AI, and Hybrid Authorship in Turkish

The proliferation of large language models capable of producing fluent Turkish prose confronts linguistic scholarship with a problem inverse to its longstanding concern: rather than analysing texts of known human provenance, scholars now face the task of determining whether a given text was authored by a human writer, generated by a machine, or composed through some combination of the two. Existing detection systems based on perplexity and burstiness suffer from opacity, English-language bias, and an inability to handle hybrid composition. The present study advances a forensic linguistic alternative grounded in the presence/absence coding framework and Jaccard similarity coefficient developed by Grant (2013) and adapted to Turkish in Kocagül (2022). Drawing on a corpus of 336 argumentative texts produced by 56 participants across six task conditions, the analysis applies 61 active linguistic features distributed across lexical, syntactic, structural, AI-specific, and data-driven categories. Three findings emerge. AI-generated Turkish texts exhibit a coherent feature profile, most distinctively elevated semicolon and comma density, sentence-initial discourse markers, recurrent ‑Arak converb chaining, and formulaic closing vocabulary. The within-author Jaccard baseline (0.41) substantially exceeds all human-to-machine comparisons (0.22-0.30), while the AI-to-hybrid pairing reaches 0.52, confirming the persistence of the AI lexical profile across editorial intervention. In edited hybrid texts, 41 of 61 features (67%) remain shifted from the writers’ baseline toward the AI baseline, indicating that lexical and discourse-organisational features pass below the editor’s threshold of attention. The findings characterise an AI-typical linguistic profile in argumentative Turkish and contribute a Turkish-language case to a corpus linguistic and forensic literature whose principal benchmarks have until recently excluded the language.