🤖 Political Preferences of Neural Networks: A nordan.ai Study
nordan.ai revealed the political sympathies of 23 models (Anthropic, OpenAI, Google, xAI, DeepSeek, etc.) by testing their "raw weights" via raw API without system prompts. The models tended to lean toward the political center.
🌍 This proves the importance of isolating architectural bias from safety layers (RLHF). The use of constrained decoding allows for bypassing standard model refusals to answer political questions.
👤 AI "opinion" depends heavily on the operating mode: for example, enabling reasoning in Kimi K2.6 changed 23 out of 35 answers.
Source 1: https://www.nordan.ai/research/which-swedish-party-do-llms-vote-for Source 2: https://huggingface.co/datasets/nordan-ai/valkompass-2026-llms