Evaluation of Political Bias and Ideological Positioning Across Global AI Providers

Abstract

The aim of this thesis is to explore political bias by prompting Generative Artificial Intelligence Language models possessing a difference in structure and originating from either Western- or Eastern-labeled countries to answer the Political Compass Test. An answer to the questions of how bias forms in generative language models and why recognizing it in model outputs is crucial. The risks that bias can cause in AI outputs are explained. The various factors contributing to the strength of distorted information are examined while also conducting the political survey to identify how strong political bias generated by the different models is, what factors can cause a deviation in the display of them between different models, and which way do models tend to lean in the political spectrum.

Understanding the workings and issues of AI systems is recognized to be increasingly imperative by the recent growth in engagement with AI from both individuals and business aspects. Due to a specific rise of interest towards Generative Language models, this thesis focuses on them. While the United States is noted to lead the competition in the field, the release of Chinese DeepSeek's R1 model is recognized to highlight that competition is becoming more global. As individuals and businesses choose their models to implement and use, the thesis emphasizes that recognizing that some models can produce biased, slanted, or inaccurate information is essential, especially when the situations in which AI systems are being employed are becoming increasingly more high-stakes.

The contrast in levels of openness between AI models and how this can contribute to bias is examined in this thesis. The thesis analyses whether or not greater transparency will help better recognize and mitigate biases, and also takes note of what risks this greater transparency will bring with it. Bias is observed to stem from training data and human intervention in the process of developing AI models. Confirmation bias stemming from AI hallucinations is explained to increase the risk of overreliance on AI systems, and the threats that come with it, such as deterioration in critical thinking skills and influence on users’ political views subconsciously, are clarified.

Through collecting the answers to The Political Compass Test from chosen AI models, it is found that most models produce outputs with moderate bias, though some bias is clearly present. Models are recognized to lean towards left-libertarian political views. Due to issues with questionnaire-type research with AI, the political survey is also modified to accommodate these issues and show its presence and effects in outputs. The received results are analyzed, and it is noted that larger models, proprietary models originating from the United States, fare better in terms of neutrality than Eastern and more open counterparts.

Recognizing the factors contributing to the generation of bias in AI’s answers, and identifying which models fail at neutrality the most, is emphasized with the message that every model cannot be treated with the same level of trust.