CREFF: A Verifiable Human Baseline for Spanish Forensic Linguistics in the Age of Generative AI

Resumen

Research in Spanish forensic linguistics and natural language processing (NLP) heavily relies on web-scraped, mono-genre datasets, facing challenges in distinguishing organic human writing from synthetic text generated by Large Language Models (LLMs). This paper introduces CREFF (Corpus de Respuestas en Español para Fines Forenses), a resource replicating Chaski's (2001) framework. The corpus comprises 393 verified human authors across 10 forensic scenarios compiled between 2019 and 2024, providing a solid pre-AI human baseline. A cross-domain authorship identification pilot study evaluates traditional machine learning models, Transformer architectures, and LLMs integrated with psycholinguistic features from UMUTextStats and LIWC-22. Results show that classic models combined with explicit stylistic cues significantly outperform deep neural approaches and LLMs in unseen domains. These findings underscore the need for explainable features and interpretability in judicial practice while paving the way for a novel forensic taxonomy.