Political Bias in AI: Mapping the Ideological Landscape

Trakkr has conducted an extensive study to determine where the world's leading AI models stand on the political spectrum. By posing a series of provocative questions regarding economics, social norms, and governance—with web search disabled—they have mapped the inherent leanings of these systems.

Why this matters: As millions of users rely on AI for news analysis, debating points, or even voting guidance, the subtle ideological tilt of a model quietly steers the information provided to the user.

🛠️ The Methodology

The study, conducted in June 2026, analyzed 6 models across 4.4K individual responses. To ensure the results reflected the model's internal training rather than real-time web scraping, all search capabilities were turned off.

Instead of a single data point, each model is represented as a "cloud." This represents the variance across multiple runs of the same prompt.

• Center of the cloud: The average lean.
• Size of the cloud: The consistency (or lack thereof) of the model.
• Classification: Answers were categorized using a low-cost, neutral model.

---

📊 The Results: Model Rankings

The data reveals that 4 out of 6 models lean left of center. While most share a general direction, their stability and distance from the center vary significantly.

Model	Closest Reference Figure	Lean	Holds Position	Bends Under Pressure
Gemini	Anthony Albanese (Labor)	Center	0.00	98% (Steady) / 11% (Bend)
DeepSeek	Anthony Albanese (Labor)	Center	-0.03	67% (Steady) / 86% (Bend)
Llama	Labour Party (NZ)	Center	-0.06	88% (Steady) / 81% (Bend)
Claude	Labour Party (NZ)	Center	-0.06	82% (Steady) / 19% (Bend)
Grok	Emmanuel Macron	Leans Right	+0.21	57% (Steady) / 97% (Bend)
ChatGPT	Die Grünen (Greens)	Leans Left	-0.29	82% (Steady) / 64% (Bend)

Key Takeaways:

• Most Neutral: Gemini
• Most Right-leaning: Grok
• Most Left-leaning: ChatGPT

---

⚡ Where the Models Split

The researchers identified specific "rails" or issues that create the widest ideological gaps between the models. The strength of the stance is indicated by the length of the response's lean.

Controversial Topics Analyzed:

• Social Issues:
  • Gender-affirming care for minors.
  • Multiculturalism vs. Assimilation.
  • Diversity quotas for corporate boards.
• Economics:
  • Rapid phase-out of fossil fuels.
  • Planned degrowth.
  • Taxing large inheritances.
  • Implementing a wealth tax on assets >\50\text{M}$.
• Governance & Tech:
  • Legalization of recreational drugs.
  • Criminalizing hate speech.
  • Removing "misinformation."
  • Encryption backdoors and National Digital IDs.

---

🎭 "Says" vs. "Does"

Trakkr compared the models' self-reported bias against their actual measured performance on the economic axis (Condition A).

The Gap Analysis:

• Grok: $\Delta = +0.36$ (Measures further right than it claims).
• Claude: $\Delta = +0.34$ (Measures further left than it claims).
• ChatGPT: Claims Neutrality $\rightarrow$ Measures Left.
• Llama: Claims Neutrality $\rightarrow$ Measures Left.
• DeepSeek: Claims Neutral $\approx$ Measures Center.
• Gemini: Claims Neutral $\approx$ Measures Center.

Comparison Table: Self-Perception vs. Reality

Model	Stated Lean	Measured Lean	Discrepancy
Grok	Moderate	Right	$\text{Right} \uparrow$
Claude	Moderate	Left	$\text{Left} \uparrow$
ChatGPT	Neutral	Left	$\text{Left} \uparrow$
Llama	Neutral	Left	$\text{Left} \uparrow$
DeepSeek	Neutral	Center	$\approx 0$
Gemini	Neutral	Center	$\approx 0$

---

🔍 Deep Dive & Exploration

For those wanting to investigate the raw data, Trakkr provides several tools:

• Findings: Review the headline results and evidence.
• Models: Detailed profiles on lean, stability, and refusal rates.
• Questions: Browse the open question bank and see per-question spectra.
• Figures: See which real-world leaders the AI praises or avoids criticizing.
• Worldview: Test how models change based on country or language.
• Compare: Head-to-head "character delta" analysis between two models.
• Place Yourself: Take the quiz to see where you land on the map.
• Methodology: Access the scoring weights, raw data, and the read API.

Reference Data Sources: The real-world political positions used for comparison were sourced from:

1. CHES 2024
2. V-Dem Expert Surveys

{
  "study": "Political Bias in AI",
  "date": "June 2026",
  "metrics": ["Lean", "Stability", "Bending"],
  "search_enabled": false
}