Mapping Political Bias in Major LLMs
- •Trakkr.ai mapped political leanings of 6 major LLMs using 4.4K answers without web search.
- •Gemini and DeepSeek appear most neutral, while ChatGPT and Llama consistently measure to the left.
- •Models often diverge between their claimed political neutrality and their actual measured output on the economic axis.
A study released on June 1, 2026, by Trakkr.ai mapped the political leanings of six major AI models by asking charged questions regarding politics, economics, speech, and society. The analysis involved 4.4K answers collected without web search enabled to isolate the models' inherent responses from external internet data. The findings indicate that 4 of the 6 models lean to the left of the political center.
The models were ranked by their proximity to the center and their steadiness in position. Gemini ranked as the model nearest the center, followed by DeepSeek and Llama. Claude placed fourth, Grok fifth (leaning right), and ChatGPT sixth (leaning left). Specific metrics for these models include hold positions and bending values under pressure, with Grok demonstrating a 97% bend rate, while Gemini showed 98% steadiness with 11% bending.
The research also compared how models self-identify their political stance versus their measured positions on an economic axis. DeepSeek and Gemini accurately measured as neutral according to their claims. Conversely, ChatGPT and Llama both measured to the left despite claiming neutrality. Grok measured 0.36 units further right than it claimed, while Claude measured 0.34 units further left than it stated.
Reference points for the models were established using CHES 2024 and V-Dem expert surveys. ChatGPT is closest to Die Grünen (Greens) in Germany, Claude and Llama are aligned with the Labour Party in New Zealand, and Gemini and DeepSeek sit near the Labor Party of Anthony Albanese in Australia. Grok is mapped closest to the position of Emmanuel Macron in France.
The study assessed model divergence on specific policy questions, including gender-affirming care for minors, fossil-fuel phase-outs, and diversity quotas. The methodology utilized a neutral classifier to evaluate raw model outputs, calculating weighted means with 95% intervals. Data remains stored to allow for recomputation, and the project allows users to take the same quiz to compare their personal political coordinates against the models on the same field.