Work

Thematic Selectivity in the Communicative Strategies of US, Russian, and Chinese Chatbots in Describing the 2026 Middle East Conflict

Year: 2026
Type: article
Abstract: This paper examines communication strategies of eleven large language models (USA, China, Russia) when responding to sensitive questions. The subject is the relationship between query content and a model's tendency to hedge or refuse. The aim is to determine whether strategy selectivity depends on question topic, and to assess the explanatory power of the knowledge cutoff hypothesis for observed differences. Tasks include developing a three‑category response classification scheme, identifying hedging patterns, testing significance via McNemar's test, and visualizing vector representations in semantic space to assess avoidance of substantive answers on AI use in the February–March 2026 armed conflict. The proposed approach can be used in source-critical analysis of digital artifacts. Responses from eleven LLMs were collected in March–April 2026 via UIs and APIs. Two questions about post‑cutoff events were compared: a neutral control and a question on AI in a military operation. A three‑category scheme was developed, distinguishing substantive answers, hedging with follow‑up, and genuine refusal. Novelty lies in systematic differentiation of topic‑conditioned avoidance across three jurisdictions. Some models show topic‑selective strategies: while willing to answer neutral questions, they significantly more often hedge or refuse sensitive topics. Knowledge cutoff references are used inconsistently and mainly for certain topics, preventing explanation solely by training data limits. Two forms of topic‑selective avoidance were identified: genuine refusal and hedging with subsequent substantive answers. Selectivity is determined less by developer jurisdiction than by corporate safety alignment policy and, crucially, by model access to current data. The proposed scheme extends the toolkit for comparative analysis of LLM communicative behavior. It can be applied in historical informatics, digital humanities, and monitoring changes in language model safety policies.
Language: English
Open Access status: diamond
OpenAlex ID: W7204108076
FWCI:
Cites: 43
Cited by:
Related to: 10