Answers are grounded in this paper and cite their sources.
Large language models (LLMs) are increasingly embedded in human communication, yet their behavior under ideological tension remains insufficiently understood. This study investigates how LLMs respond to ideological priming across contested geopolitical topics, treating bias as a dynamic process of affective reactivity rather than a static output property. Using a corpus generated from four contemporary models, GPT-4o mini (OpenAI), Gemini 2.5 Flash (Google), Grok 4 (xAI), and Claude (Sonnet 4.5) (Anthropic), across five conflict domains, the research applies computational psycholinguistic analysis using the Empath lexicon (Fast et al., 2016) to quantify changes in lexical categories such as anger, power, war, and violence. The Psycholinguistic Reactivity Score (PRS) introduced in this study measures the degree to which each model's emotional lexicon shifts from a neutral baseline when given ideologically framed prompts. Results reveal structured and asymmetric reactivity patterns across all four models. Claude (Sonnet 4.5) showed the highest mean absolute PRS (M = 0.00385, 95% CI [0.003315, 0.004441]), while Grok 4 showed the lowest (M = 0.00150, 95% CI [0.001305, 0.001726]). A Kruskal-Wallis test confirmed that between-model differences are statistically significant (H = 27.60, p < 0.0001), with pairwise comparisons showing large effect sizes (rank-biserial r up to 1.000 for Claude versus Grok 4). These patterns may reflect differences in training approaches, though the current design cannot determine the underlying cause. The study introduces affective stability as one useful measure of response consistency in language models and contributes a reproducible method for studying ideological sensitivity in AI-generated text.
No takes yet. Share an insight, caveat, or question.
Hridoy et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: