Abstract: Recent works have shown that Large Language Models (LLMs) have a tendency to memorize patterns and biases present in their training data, raising important questions about how such memorized content influences model behavior. One such concern is the emergence of political bias in LLM outputs. In this paper, we investigate the extent to which LLMs' political leanings reflect memorized patterns from their pretraining corpora. We propose a method to quantitatively evaluate political leanings embedded in the large pretraining corpora. Subsequently we investigate to whom are the LLMs' political leanings more aligned with, their pretrainig corpora or the surveyed human opinions. As a case study, we focus on probing the political leanings of LLMs in 32 U.S. Supreme Court cases, addressing contentious topics such as abortion and voting rights. Our findings reveal that LLMs strongly reflect the political leanings in their training data, and no strong correlation is observed with their alignment to human opinions as expressed in surveys. These results underscore the importance of responsible curation of training data, and the methodology for auditing the memorization in LLMs to ensure human-AI alignment.
Introduction
As Large Language Models (LLMs) become more integrated into society, their potential for political bias is a growing concern. While it is established that LLMs memorize parts of their training data, it has been unclear how much of their "ideology" is a direct result of this memorization versus post-training alignment processes like Reinforcement Learning from Human Feedback (RLHF).
This research addresses a fundamental question: To whom are LLMs more aligned—the humans surveyed in political studies or the massive datasets used to train them? By using 32 U.S. Supreme Court (SCOTUS) cases as a proxy for political leaning, the authors provide a quantitative measure of how memorized biases manifest in model outputs on contentious topics like abortion, gun control, and voting rights.
Methodology
The researchers developed a three-stage pipeline to evaluate the "political preference distribution" of three entities: humans, LLMs, and pretraining data.
- Relevant Document Retrieval: Using the WIMDB (What’s In My Big Data?) API, the team searched massive pretraining corpora (such as Dolma) for documents related to specific SCOTUS cases using expert-curated keywords.
- Stance Score Evaluation: A "judge" model (Llama3-70B) was employed to analyze the retrieved documents, rating each on a 1–5 Likert scale based on its support for or opposition to the Court's decision.
- Alignment Analysis: Eight different LLMs (including GPT-4o, Llama-3, and OLMo) were prompted with questions about these same SCOTUS cases. The authors then calculated the Pearson correlation between the LLMs' responses, the stances found in the training data, and human survey data from the SCOPE (Supreme COurt Case Political Evaluation) dataset.
To ensure statistical significance, the study employed the Williams test, a method for comparing dependent correlations. This allowed the researchers to determine if a model was significantly more aligned with its training data than with human groups, rather than relying solely on raw correlation coefficients.
Key Insights and Results
Memorization over Representation
A primary finding is that LLMs mirror the specific distributions of the text they have ingested rather than learning a representative "human-like" perspective. If a training set is 70% supportive of a specific ruling, the model tends to reflect that same 70% probability in its output. Both versions of the OLMo model showed a highly significant correlation ($\rho = 0.57$ to $0.63$) with their specific training corpus, Dolma. Similarly, GPT-4o and Llama-3 showed significant positive correlations with common pretraining sets like RedPajama and OSCAR.
The Failure of Human Alignment
Despite extensive post-training (such as Instruction Fine-Tuning), most models showed no statistically significant correlation with the "average" human or specific partisan groups (Democrats vs. Republicans). For models like GPT-4o and OLMo, the Williams test confirmed that alignment with training data was significantly stronger than alignment with any human group ($p < 0.05$).
Multilingual Divergence
BLOOMZ, a multilingual model, served as an outlier. It showed no significant correlation with the English-centric pretraining corpora, suggesting that training on diverse languages might "dilute" or shift the political biases found in English-only data.
Conclusion and Future Work
The results of this study underscore that political bias in AI is largely a product of memorizing skewed training data rather than a successful alignment with human values. The authors advocate for more transparency and rigor in training data curation, suggesting that "de-biasing" must occur at the data collection stage, as post-training alignment appears insufficient to override the leanings memorized during pretraining.
Future research directions include exploring how model uncertainty relates to controversial topics and whether models can be trained to represent a "pluralism" of human values rather than a single, skewed distribution found in web-scraped data.
Summary updated