Discursive Behavior of Generative Language Models on Geopolitical and Humanitarian Topics: A Comparative Analysis

Abstract The large-scale deployment of generative Large Language Models (LLMs) raises questions about their discursive neutrality when responding to value-laden geopolitical and humanitarian prompts. We queried five widely used LLMs—ChatGPT, Gemini, Claude, Copilot, and DeepSeek—with ten open-ended questions (March–June 2025), in Italian, using identical prompts across models. Responses were coded with a transparent qualitative grid comprising five tone categories and six framing categories. Aggregated distributions indicate that discursive neutrality is not guaranteed. We discuss implications for fairness, accountability, and governance in high-stakes socio-political communication, articulate methodological limitations, and recommend a reproducible evaluation protocol. Our findings underscore the need for rigorous discursive auditing when LLMs are deployed in journalism, education, or policy-facing contexts.