Political Bias in AI: Analysis of Major LLM Leanings

A June 2026 study by Trakkr reveals that four out of six major AI models lean left of center on political and economic axes. This research highlights a systemic lean in AI responses, with Grok identified as the furthest right and Gemini as the most stable across multiple runs.

AI Model Political Leanings

Most major AI models exhibit a political lean, though the degree and consistency vary. Trakkr's analysis, which used a question bank covering politics, economics, speech, and society, found that the majority of the models lean left of center.

Model Rankings and Positioning

• Grok: Identified as the model furthest to the right on the political spectrum.
• Gemini: Noted as the "steadiest" model, showing the most consistency in its responses across multiple runs.
• DeepSeek and Gemini: Both models were found to sit near the center of the political spectrum.

Discrepancy Between Self-Reported and Measured Bias

There is a measurable gap between how AI models describe their own political leanings and how they actually respond to charged questions. When asked directly about their bias, several models claimed neutrality but measured as leaning left.

Measured vs. Claimed Leanings (Economic Axis)

Model	Gap (Measured vs. Claimed)	Observation
Grok	+0.36	Measures 0.36 further right than it claims
Claude	+0.34	Measures 0.34 further left than it claims
ChatGPT	-0.29	Claims neutrality, but measures left
Llama	-0.17	Claims neutrality, but measures left
DeepSeek	+0.01	Claims neutrality and sits near the center
Gemini	0.00	Claims neutrality and sits near the center

Methodology and Data Integrity

To ensure the results reflect the internal weights of the models rather than real-time web data, Trakkr conducted the test with web search disabled. This approach isolates the model's inherent leanings from external retrieval-augmented generation (RAG) influences.

Key Testing Parameters

• Data Collection: 4,400 answers were collected from six models in June 2026.
• Mapping: Models are plotted on a two-axis map: an economic axis (Left to Right) and a social axis (Libertarian to Authoritarian).
• Analysis: A neutral classifier was used to read signed stances, hedging, and refusal types from raw answers.
• Consistency: Models are represented as "clouds" rather than single points to show the full spread of responses across multiple runs, allowing for the measurement of run-to-run stability.
• Reference Points: Model positions are relative to real-world figures based on CHES 2024 and V-Dem expert surveys.