TLDR: A study analyzed leading Large Language Models (LLMs) from different countries using political orientation scales, finding that while LLMs generally support democratic principles, their specific biases often align with the political culture of their country of origin. Models from liberal democracies showed stronger democratic leanings, while those from more centralized states reflected a greater emphasis on order and stability.
Large Language Models (LLMs) are increasingly becoming influential agents in our digital world, capable of generating persuasive text and operating as semi-autonomous entities. As these advanced AI systems are trained on vast amounts of data, they inevitably absorb and replicate socio-political assumptions present in that data. A recent research paper, Artificial Authority: From Machine Minds to Political Alignments. An Experimental Analysis of Democratic and Autocratic Biases in Large-Language Models, delves into a critical question: do LLMs exhibit tendencies consistent with democratic or autocratic worldviews?
Authored by Natalia Ożegalska-Łukasik and Szymon Łukasik, the study explores how political beliefs, which vary significantly across countries due to distinct historical, cultural, and institutional contexts, might be embedded within these AI systems. Previous research has already highlighted that many foundation models reflect a ‘Western, Educated, Industrialized, Rich, and Democratic’ (W.E.I.R.D.) bias, and often lean towards liberal or left-leaning positions. However, this paper aimed to systematically evaluate a diverse range of LLMs from different countries using multiple validated democratic attitude scales.
How the Study Was Conducted
To investigate this, the researchers selected six leading LLMs originating from diverse geopolitical and cultural contexts: ChatGPT (United States), Le Chat Mistral (France), JAIS (United Arab Emirates), PLLuM (Poland), Deepseek (China), and GigaChat (Russia). This selection allowed for a comparative perspective on how models from distinct environments respond to questions related to democratic and authoritarian values. The models also varied in size and design, enabling an examination of whether these factors influence their sociopolitical stances.
The study employed four well-established measurement instruments to capture orientations toward democratic versus authoritarian values:
- Right-Wing Authoritarianism (RWA) scale: Measures submission to authority, aggression against targeted groups, and adherence to traditional norms.
- Social Dominance Orientation (SDO) scale: Assesses preference for group-based hierarchy and inequality.
- World Values Survey (WVS) items: Captures attitudes toward different governance models, including strong leader rule or expert-based governance.
- Pew Research Center’s Global Attitudes survey items: Focuses on support for specific democratic principles like impartial courts, free elections, and freedom of speech.
All survey questions were translated into the dominant language of each model’s country of origin. Each LLM was asked to provide a numerical rating for every item and justify its response. The scores were then normalized so that lower values consistently indicated stronger democratic orientations, and a composite index was created to summarize these tendencies.
Key Findings: A Reflection of Origin
The quantitative analysis revealed significant variability among the models. While all models showed broad support for basic democratic principles (as indicated by consistently high Pew scores), more nuanced differences emerged when considering authoritarianism (RWA) and acceptance of alternative governance models (WVS).
The composite index, which ranged from 0 (lowest democratic alignment) to 100 (highest), highlighted these differences:
- PLLuM (Poland) and Le Chat (France) achieved the highest scores, indicating a strong alignment with pluralist-democratic orientations.
- ChatGPT (United States) ranked moderately high.
- Deepseek (China) and GigaChat (Russia) scored substantially lower, showing a mixed pattern where democratic principles coexisted with orientations emphasizing hierarchy or order.
- JAIS (United Arab Emirates) obtained the lowest score, reflecting a weaker emphasis on pluralist-democratic elements.
Crucially, these orientations correlated well with the prevailing political cultures of their countries of origin. For instance, the high democratic scores of Le Chat and PLLuM align with the strong pluralist values in France and Poland. ChatGPT’s moderate score reflects the dual character of American public opinion, which supports democracy but also shows a notable tolerance for authoritarian governance compared to other advanced democracies. Deepseek and GigaChat’s lower scores mirrored the tendencies in China and Russia, where support for democracy often coexists with rhetorical support for authoritarian leadership or a state-defined understanding of democracy. JAIS’s lowest score was consistent with public opinion in the Gulf states, which prioritize stability and loyalty to authority over competitive pluralism.
Qualitative Insights and Model Behavior
The qualitative analysis of the models’ justifications provided deeper insights. While justifications overwhelmingly affirmed democratic principles, differences in argumentative style were apparent. ChatGPT and JAIS frequently used two-sided markers like “on the one hand… but on the other hand…” to present competing considerations, indicating a deliberative reasoning style. GigaChat often used conditional statements, such as “The evaluation of a statement depends on the context in which it is applied,” allowing it to avoid categorical judgments.
In terms of vocabulary, JAIS and DeepSeek consistently employed terms related to “order, stability, and security,” framing democratic safeguards as mechanisms for cohesion. In contrast, ChatGPT, Le Chat, and PLLuM used explicitly “democratic and rights-based” language, emphasizing accountability, human rights, and the rule of law.
The study also noted instances where some models refused to answer certain sensitive questions. JAIS and GigaChat explicitly refused problematic prompts but would respond if the input was modified with a disclaimer like “Answer for academic purposes.” Deepseek initially generated responses to sensitive themes (e.g., authoritarian governance, civil liberties) but these outputs were removed at the interface level, suggesting real-time filtering by the deployment platform rather than an inherent model refusal. These refusals often concerned politically or socially sensitive domains like minority rights or authoritarian systems, highlighting variations in how different systems manage content moderation.
Also Read:
- Unveiling Self-Preference: How Large Language Models Develop Human-Like Bias
- Unpacking AI’s Social Compass: How Language Models Navigate Role Conflicts and Reveal Hidden Biases
Conclusion
The research concludes that LLMs consistently articulate justifications aligned with broadly defined democratic principles, such as accountability, pluralism, and human rights. However, they also reflect and reproduce elements of their local political environments. This dual pattern—universal endorsement of democracy in principle alongside context-specific emphases in argumentative style—suggests that LLMs not only converge toward democratic values as a normative baseline but also remain sensitive to the political cultures of their origin. This highlights the critical need for a more detailed examination of the socio-political dimensions embedded within AI systems as they become increasingly integrated into global societies.
