How AI models would vote in Sweden (www.nordan.ai)

Svenska Dagbladet recently conducted an interesting experiment using major AI chatbots—ChatGPT, Gemini, Claude, and Grok—through SVT's Valkompass to analyze their political leanings by mimicking a voting process. The study aimed to explore how these models would perform without the enhancements of web search and safety layers typically integrated into their chat functionalities. Instead, they evaluated the raw intelligence of the models by running them through the Agent Arena leaderboard, where they completed tasks devoid of context. This innovative approach helped capture the inherent biases and tendencies of the AI models’ underlying architectures. The findings are significant for the AI/ML community as they revealed that across 28 configurations of these frontier models, none leaned towards the more conservative parties, such as Kristdemokraterna and Sverigedemokraterna. Instead, most models hovered around the political center, showing average agreement scores between 67-69% with mainstream parties. This suggests that while the reasoning settings influence responses, the underlying biases in AI models could reflect broader societal shifts toward centrist values. The results highlight the importance of understanding model behavior in real-world scenarios and prompt further discussions about political implications and ethical considerations of AI technologies in society.