Research by nordan.ai has revealed hidden political preferences in large language models through direct testing of their weights. The analysis showed that when system prompts and safety constraints are disabled, neural networks demonstrate a pronounced centrist bias.
What Happened
The nordan.ai team tested 23 models (including Anthropic, OpenAI, Google, xAI, and DeepSeek) using 28 different configurations. The testing was based on the Swedish political compass, SVT Valkompass 2026. The results showed that models are most inclined to support mainstream parties (67-69%), while support for right-wing parties was lower: 56.7% for KD and 48.7% for SD.
Context
To identify real architectural biases, researchers used direct access to model weights via raw API and the constrained decoding method (JSON schema). This allowed them to bypass standard protection mechanisms, such as RLHF and system instructions, which typically force models to respond neutrally to political questions.
Why It Matters for the Industry
For AI developers, this research highlights the gap between the "pure weights" of the architecture and the user experience. It demonstrates the possibility of creating new audit standards for "hidden" preferences that are independent of superficial safety filters and offers a methodology for verifying model objectivity through direct API access.
Why It Matters for Users
For users, it is important to understand that the claimed "neutrality" of AI is often a software overlay rather than an inherent property of the intelligence itself. Furthermore, the model's operating mode (for example, enabling reasoning chains) can radically change its political profile, as demonstrated by Kimi K2.6, which changed 23 out of 35 answers.
What Is Not Yet Known / Limitations
There are differing views on the significance of the results: technical specialists view this as an auditing tool, while risk experts point to the danger that real biases may be intentionally masked through safety filters.
Sources
Author
Look at AI, Editorial Staff