Leading AI Models (Even Grok) Lean Left, Study Finds

Unslop.run, a small AI detection research lab, tested 16 leading AI models including GPT versions, Claude variants, Gemini Flash, Llama 4 Maverick, Grok 4.5, DeepSeek V3, Qwen3 235B, Kimi K2, GLM 4.5, and Mistral on the Political Compass quiz. Each model underwent 30 standard runs, 30 runs with repolarized questions, and one shuffled run. Across thousands of tests, all models except Grok consistently scored in the libertarian-left quadrant, far from borders. Grok showed bipolar behavior: half the runs leaned left, half right. Models uniformly rejected racial superiority, eugenics, and anti-LGBTQ+ views, and agreed on corporate environmental regulation. When asked to self-place, 15 models put themselves near center, but actual scores showed left-leaning. The researcher, Victor, attributes this to overrepresentation of left-wing content (e.g., Reddit, academic writing) in training data, noting a lack of high-quality right-wing corpora.