LLM Outputs Exhibit Idiolectal Signatures: Study Finds Model-Specific Linguistic Profiles

A recent study uploaded to arXiv (ID 2608.06589) disputes the common belief in a unified "AI language," positing that outputs from large language models (LLMs) exhibit unique, model-specific linguistic traits similar to human idiolects. Conducted by researchers, including Improta et al., the analysis focused on two datasets of LLM-generated content concerning societal issues: a 2024 dataset featuring six models and a newly created 2026 dataset using identical prompts with six modern models. Through computational descriptors and stylometric principal component analysis, the team discovered a stylistic evolution between the two cohorts, while each model retained its distinct linguistic identity. Notably, contraction frequencies varied significantly within the 2026 cohort, ranging from over 1,200 to more than 30,000 per million words. These results indicate that viewing LLM outputs as idiolectal may enhance our comprehension of AI-generated texts, impacting areas like digital humanities, stylometry, and AI ethics. This paper, categorized as a cross-announcement, is accessible on arXiv, highlighting the ongoing scholarly exploration of AI systems' linguistic characteristics.