💬 NLP

What Is The Political Content in LLMs' Pre- and Post-Training Data?

This paper investigates the political content in LLM training data and finds that pre-training corpora are systematically skewed toward left-leaning perspectives, with these biases persisting through post-training stages and strongly correlating with the resulting model's political stances, thereby highlighting the critical role of data composition in shaping model behavior.

Original authors: Tanise Ceron, Dmitry Nikolaev, Dominik Stammbach, Debora Nozza

Published 2026-04-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tanise Ceron, Dmitry Nikolaev, Dominik Stammbach, Debora Nozza

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are hiring a new assistant to help you write emails, answer questions, and give advice. You don't just hand them a blank notebook; you give them a massive library of books, articles, and blog posts to read before they start working. This is exactly how Large Language Models (LLMs) like the ones powering chatbots are built. They "read" terabytes of internet data to learn how to speak and think.

This paper asks a simple but crucial question: What kind of library did we give these AI assistants, and does that library make them biased?

Here is the breakdown of their findings, explained with some everyday analogies.

1. The "Left-Leaning" Library

The researchers looked at the "books" (data) used to train three different famous AI models (OLMo, Falcon, and Pythia). They wanted to know: Is the library balanced, or is it stacked?

The Finding: The library is heavily stacked toward left-leaning viewpoints.
The Analogy: Imagine you are trying to learn about politics by reading a stack of newspapers. If 80% of the papers are from The Guardian or HuffPost (left-leaning) and only 10% are from National Review or The Telegraph (right-leaning), you are going to learn a very specific version of the world.
The Numbers: In the data they analyzed, there were 2.3 to 12 times more left-leaning documents than right-leaning ones. It's like a diet where you only eat one flavor of ice cream; eventually, you start thinking that's the only flavor that exists.

2. Where Do the Books Come From? (The Source)

The researchers also looked at where these documents came from.

The Finding: Left-leaning texts mostly came from established news outlets (like The New York Times or The Washington Post). Right-leaning texts mostly came from blogs, personal websites, and fringe communities (like Townhall or FreeRepublic).
The Analogy: Think of the left-leaning data as a formal dinner party hosted by professional journalists. The right-leaning data is more like a rowdy backyard BBQ with passionate individuals shouting their opinions. The AI "ate" both, but the formal dinner party provided much more food.

3. The "Echo Chamber" Effect

One of the big questions was: If we use different libraries for different AI models, will they have different biases?

The Finding: Surprisingly, no. Even though the researchers used different data sources and different filtering methods for different models, the political bias was almost identical.
The Analogy: Imagine three different chefs (different AI models) using three different recipes (different datasets) to make soup. You'd expect them to taste different. But instead, they all tasted exactly the same because they all used the same main ingredient: a massive, unfiltered dump of the internet that happens to be naturally left-leaning. The "filtering" didn't change the flavor much.

4. The "First Bite" vs. The "Dessert"

AI training happens in stages:

Pre-training: The model reads the massive library (the main course).
Post-training: Humans tweak the model to be more helpful and polite (the dessert).

The Finding: The political bias is baked into the main course (pre-training). By the time the humans try to tweak the model with the "dessert" (post-training), the bias is already there. The dessert didn't change the flavor; it just made the AI speak more politely about its existing opinions.
The Analogy: If you teach a child that "the sky is green" when they are 2 years old (pre-training), and then at age 10 you tell them to "be polite and helpful" (post-training), they will still think the sky is green. They will just be very polite about it. The bias was established early and stuck.

5. The Connection: Data = Behavior

Finally, the researchers checked if the AI's actual answers matched the books it read.

The Finding: There was a very strong match (90% correlation). If the training data said "Climate Change is urgent," the AI said "Climate Change is urgent." If the data said "Immigration should be restricted," the AI was more likely to agree with restrictions (though the data was mostly against restrictions, so the AI was mostly against them too).
The Analogy: The AI is a parrot. It doesn't have its own opinions; it just repeats the patterns it heard most often in the room. If the room was full of people shouting left-wing ideas, the parrot will sound left-wing.

The Big Takeaway

The paper concludes that we cannot fix AI bias just by tweaking the model later. The bias is in the ingredients.

If we want AI to be truly balanced and fair, we need to fix the library before the AI starts reading. We need to ensure the "books" we feed it represent a true mix of voices, not just the loudest or most common ones on the internet.

In short: The AI isn't "evil" or "biased" by choice. It's just a mirror reflecting the skewed library we gave it. If the library is unbalanced, the reflection will be too.

1. Problem Statement

Large Language Models (LLMs) are increasingly used as sources of factual information and political commentary, yet they are known to generate politically biased text. While prior research has extensively documented that LLMs exhibit systematic left-leaning biases, the origins of these biases remain poorly understood. Most existing studies adopt a model-centric approach, evaluating model outputs without analyzing the underlying training data. This limits the ability to design effective mitigation strategies. The authors argue that a data-centric perspective is essential to understand how biases arise, persist, and are reinforced across different training stages (pre-training vs. post-training).

2. Methodology

The study employs a large-scale, data-centric analysis of open-source LLMs and their associated datasets.

Models and Datasets Analyzed:
- Models: OLMO2-13B, Falcon2-11B, and Pythia-12B.
- Datasets: The authors selected models with minimal training data overlap to ensure independence. They analyzed:
  - Pre-training: DOLMA, DOLMINO (OLMO2), REFINEDWEB (Falcon2), and THE PILE (Pythia).
  - Post-training: SFT-MIX (Supervised Fine-Tuning) and DPO-MIX (Direct Preference Optimization).
Sampling Strategy:
- Due to computational constraints, the authors used reservoir sampling to extract representative subsets (ranging from ~100k to 300k documents) from massive pre-training corpora.
- Non-English documents were filtered out using a fastText-based classifier.
Political Leaning Classification:
- Classifier: A zero-shot classification pipeline using LLaMA-3.1-70B-Instruct was employed to categorize documents into Left-leaning, Right-leaning, or Neutral.
- Validation: The classifier was validated against NEWSLEAN, a proprietary dataset of 4,434 articles from 94 US news outlets manually annotated by experts. The classifier achieved a precision of 0.72 for right-leaning and 0.65 for left-leaning documents.
Stance Detection:
- Model Stances: Evaluated using the ProbVAA dataset (239 statements across 8 policy vectors). Models generated 30 responses per statement variant to calculate a stance score $S = (A-D)/(A+D)$ .
- Document Stances: A zero-shot classifier (LLaMA-3.1-70B) was used to detect support/opposition to the same 8 policy vectors within the training documents.
Analysis Metrics:
- Correlation: Pearson correlation coefficients ( $r$ ) were calculated between the political stances in training data and the resulting model stances.
- Source Domain Analysis: Metadata URLs were extracted to analyze the origin of political content (e.g., blogs vs. news outlets).
- Topic Modeling: BERTopic was used to cluster documents and analyze framing differences between left and right-leaning texts.

3. Key Contributions

Systematic Quantification: Provided the first large-scale quantification of political leaning distributions in both pre-training and post-training datasets across multiple open-source models.
Data Imbalance Discovery: Demonstrated that left-leaning content consistently dominates training data, outnumbering right-leaning content by factors ranging from 2.3x to 12x.
Cross-Model Similarity: Showed that despite different data curation strategies and sources, pre-training datasets exhibit highly similar political distributions (high Rank-Biased Overlap).
Training Stage Analysis: Revealed that political biases are established during pre-training and persist through post-training stages (SFT and DPO), even though post-training data contains significantly less political content.
Source Domain Insights: Identified that right-leaning content primarily originates from blogs and niche online communities, whereas left-leaning content is more frequently sourced from established news outlets.

4. Key Results

R1 (Imbalance): Confirmed. Left-leaning documents consistently outnumber right-leaning ones. Pre-training corpora contain 2.5 to 4 times more politically engaged material than post-training data.
R2 (Variance): Refuted. Despite different curation strategies (e.g., DOLMA vs. REFINEDWEB vs. THE PILE), the political distributions are highly correlated ( $r \approx 0.97-0.99$ ), suggesting curation strategies do not significantly alter aggregate political properties.
R3 (Correlation): Strongly Confirmed. There is a strong correlation ( $r = 0.87$ ) between the political stances in the training data and the stances expressed by the models across policy issues. This suggests model behavior mirrors the data distribution.
R4 (Training Stages): Partially Confirmed. Base models already encode political biases. Post-training (SFT/DPO) does not fundamentally alter the direction of the bias, though it may increase consistency. The bias is primarily established in pre-training.
Framing Differences:
- Climate: Right-leaning texts focus on economic stability and deregulation; Left-leaning texts emphasize urgency and systemic transformation.
- Religion: Right-leaning texts reinforce biblical authority; Left-leaning texts critique institutional abuses.
- Animal Rights: Right-leaning texts emphasize personal responsibility; Left-leaning texts focus on structural harms.

5. Significance and Implications

Root Cause Identification: The study shifts the focus from model architecture to data composition as the primary driver of political bias. It suggests that simply adjusting model outputs (post-training alignment) is insufficient if the foundational pre-training data is skewed.
Mitigation Strategies: The findings suggest that effective bias mitigation requires pre-filtering pre-training data to remove heavily politicized content or rebalancing the dataset, rather than relying solely on post-training alignment techniques.
Transparency and Accountability: The authors argue for greater dataset transparency. Policymakers and developers need detailed documentation of data sources to understand and regulate the political leanings of AI systems.
Societal Impact: Since LLMs act as information intermediaries, their systematic left-leaning bias could inadvertently influence public opinion, voting behavior, and democratic discourse, potentially deepening polarization or under-representing specific viewpoints.

Conclusion

The paper concludes that political biases in LLMs are not an artifact of model architecture or post-training alignment alone but are deeply rooted in the systematic skew of pre-training data. The strong correlation between data distribution and model behavior underscores the need for data-centric interventions to create more impartial and trustworthy AI systems.