Bias in the Media? How LLMs (GPT-4o, Claude, PyABSA) Strengthen Tonal Analysis:

Is it possible to detect bias in journalistic media using artificial intelligence? We present an innovative pipeline to analyze 100,000 articles from Deutsche Welle (DW), evaluating whether certain entities (people, countries, organizations) receive different tonal treatment. The solution combined three key stages:

• Scalable search engine: Automatic identification of relevant entities within corpus subsets.
• Human benchmark: Manual analysis of a sample to validate the model’s findings.
• Multimodal pipeline: Integration of GPT-4o, Claude, and PyABSA to analyze sentiment directed at specific entities.

The most striking result: a model voting system (where the majority decides the tone) significantly outperformed the accuracy of any individual model. This methodology not only provides transparency in media analysis, but also raises critical questions about the role of LLMs in auditing journalistic bias.