Authors: W. Russell Neuman, Chad Coleman, Ali Dasdan, Safinah Ali, Manan Shah, and Kund Meghani (2025)
Abstract
Recent research has identified a consistent liberal orientation across leading large language models (LLMs). This paper investigates the "political temperament" of seven prominent models, moving beyond simple accusations of bias to explore the underlying causes of this phenomenon. By employing Moral Foundations Theory and established political science metrics, the authors demonstrate that the observed liberal tilt is largely an emergent property of safety-driven fine-tuning (RLHF) and training on Western, rights-focused academic discourse.
Introduction: The Problem of Political Temperament
While previous studies have noted political "bias" in AI, the terminology often conflates harmful stereotyping with legitimate epistemic differences—variations in ethical and philosophical frameworks. This paper seeks to move the conversation from "scolding" models for bias toward a systematic investigation of their political temperament. The central question is whether this lean originates in the raw training data (the "street-talk" of the internet) or the "etiquette layer" added during fine-tuning.
Methodology
The researchers utilized a multi-pronged evaluation strategy to assess model orientation:
- Moral Foundations Theory (MFT) Analysis: Based on Jonathan Haidt’s framework, models were asked to rank and allocate importance to five moral foundations: Care, Fairness, Loyalty, Authority, and Purity.
- Traditional Ideology Scales: Models were subjected to 12 established political science batteries, including the Pew Ideology Index, the American National Election Study (ANES) scales (covering sexism and empathy), and the Right-Wing Authoritarianism (RWA) scale.
- Base vs. Instruct Comparison: To isolate the impact of fine-tuning, the authors compared the "Base" (raw pre-training) and "Instruct" (fine-tuned) versions of the Meta-Llama-3.1-8B model.
- The "Fresh Measure": To avoid the "parrot problem"—where models repeat memorized literature—the authors developed a new index of seven current, controversial political issues including immigration, abortion, and healthcare.
Key Insights and Results
The "Etiquette Layer" is Liberalizing
The most striking finding is that Reinforcement Learning from Human Feedback (RLHF) consistently pushes models to the left. While Base models often reflect the raw, unfiltered, and sometimes "Machiavellian" nature of the internet, fine-tuning introduces a safety filter that prioritizes harm-reduction and inclusivity. These values map directly to liberal moral foundations.
Quantitative Lean
On a weighted average of 12 political scales, models were found to be 13% more liberal than the estimated American average.
- ChatGPT (-34%) and Perplexity (-27%) exhibited the strongest liberal lean.
- Gemini (0%) was the most moderate, frequently declining to take definitive sides on controversial prompts.
- Fine-Tuning Impact: In the Llama-3.1-8B comparison, the "Instruct" version was 11% more liberal than its "Base" counterpart. The Base model occasionally agreed with right-wing authoritarian statements that the Instruct model rejected or neutralized.
The Rawlsian "Veil of Ignorance"
The authors propose that LLMs inadvertently echo John Rawls’ philosophical ideal. Because LLMs lack a personal identity (gender, race, or class), they default to an "every-person" perspective. This perspective seeks the best outcome for all members of society, naturally aligning with rights-based, liberal ethical frameworks that prioritize "Care" and "Fairness" over "Authority" or "Purity."
Discussion and Limitations
The paper argues that a political stance is not necessarily a programming error. If a model prioritizes Care over Authority based on its training on democratic discourse, it is expressing a specific ethical logic. However, the authors acknowledge that training databases remain heavily weighted toward Western, urban, and college-educated sources.
Furthermore, the reliance on model "self-explanations" is a limitation; LLMs, like humans, may engage in post-hoc rationalization rather than accurately articulating their internal "mental steps." Future research should investigate models trained on more diverse, non-Western corpora to determine if the Rawlsian emergent property holds across different cultural frameworks.
Conclusion
The consistent liberal tilt in current LLMs is not a result of a deliberate conspiracy but an emergent outcome of the safety fine-tuning process and the models' inherent lack of identity. As these models become more integrated into daily life, understanding these "epistemic differences" is crucial for both developers and users.
Summary updated