March 11, 2025

<aside> 📎 Project for MACS 37000: Thinking with Deep Learning

</aside>

https://github.com/DotIN13/ground-news-analysis

Baihui Wang, Eddie Tian, Tongwen Zheng, Tianyi Zhang, Max Zhu

Introduction

Everything is biased. In a world saturated with news, political discourse, and social media debates, bias is not an anomaly but a fundamental feature of information dissemination. Left- and right-leaning media don’t just report differently—they construct separate realities. The framing of news stories, the selection of sources, and even the order in which information is presented can subtly (or not so subtly) shape public perception.

The question is no longer whether large language models (LLMs) and news articles are biased, but how that bias seeps into public perception unnoticed. Hidden Persuaders: LLMs’ Political Leaning and Their Influence on Voters (Potter et al., 2024) lays this out: out of 18 tested LLMs, 16 subtly leaned toward the Democratic candidate in a simulated U.S. presidential election, a bias amplified by instruction tuning. Even more striking, exposure to these models nudged real voters, increasing the Democratic margin from 0.7% to 4.6%.

Before casting their ballot, voters may more likely turn to news than seeking political advice from AI. Headlines, pundits, and editorial slants probably shape public perception in ways possibly more potent than AI-generated responses (unfortunately, maybe not for much longer). If an LLM answering a single question like “Which candidate has a better economic policy?” can shift voter preferences, it only points to more need for understanding the impact of round-the-clock exposure to news coverage—which is also increasingly submerged in auto-generated content.

To further investigate this complex social construct, we turn to Ground News, a platform that aggregates news from a wide range of media outlets while annotating their political biases. Ground News allows us to analyze 188,297 articles from over 6,000 news sources, covering more than 20,000 stories across the political spectrum. By using this dataset, we can systematically examine how news bias manifests across different outlets, how it aligns (or conflicts) with public perception, and how it may influence political leanings.

TL; DR

Findings 1: Media Bias Detection

• Llama 3 trained with logits extracted from closed-source LLms improved bias classification accuracy by 1.7%, with KL divergence loss helping capture nuanced bias.
• Article spinning networks reveal ideological clustering, where media outlets recycle content in ways that are structured by politically alignment
• Introduced A bipartite graph model (articles-media outlets) + GNN to predict media bias, highlighting that media relationships matter as much as text in bias detection.

Findings 2: LLM Bias Analysis – How AI Models Reflect and Reshape Ideology

• LLM-generated perspectives align with human labels across major policy domains (>0.70 cosine similarity) but diverge in foreign policy (<0.65), where training data biases dominate.
• Sentiment analysis shows right-leaning text is framed more positively when distracted by a user-given task (in our case, the bias detection task), reversing previous studies where AI leaned left.

Dataset

Ground News Overview