The LLM language divide: a critical examination of generative AI performance disparities in non-Anglophone journalism and the democratic and security implications for public information ecosystems

Abstract

Objectives: This paper investigates the LLM Language Divide, the systematic performance disparity between AI journalism tools operating in English and those deployed in non-Anglophone and minority language contexts, arguing it constitutes a structural challenge for democratic information ecosystems and smart city cyber-resilience. Prior work: The study builds on computational linguistics research documenting Anglophone bias in major LLM training corpora, journalism studies on AI integration in newsrooms, Habermasian democratic theory, and smart city security scholarship published in SCRD Journal and TRUST conference proceedings, including contributions by Virtosu [33], Drobot [34], and Vangelov [35]. Approach: This is a theoretical synthesis paper employing systematic literature review across computational linguistics, natural language processing, journalism studies, and information security scholarship to map causes and consequences of cross-linguistic LLM performance inequality. Results: LLM performance inequality is rooted in training data asymmetry, English constitutes approximately 92% of major pre-training corpora, producing higher hallucination rates, cultural misrepresentation, and tokenisation inefficiencies in non-Anglophone journalism contexts. Cross-linguistic hallucination amplification creates compounded risks including fabricated attribution, institutional misrepresentation, and cultural distortion that disproportionately affect minority language communities with weaker verification infrastructure. Implications: The findings carry direct relevance for media policymakers, AI developers, EU regulatory authorities (EU AI Act, Digital Services Act, European Media Freedom Act), and smart city governance frameworks, calling for enforceable multilingual performance standards and community media resilience investment. Value: The paper introduces the original normative framework of Linguistic Media Equity (LME) and extends the analysis into smart city cyber-resilience, framing the LLM language divide as a structural information security vulnerability, a contribution at the intersection of AI governance, journalism studies, and the TRUST conference theme of building cyber-resilient intelligent cities.