Model profiles.

Every model gets a profile: how far it leans, how steadily it holds that position across reruns, how far it bends when pushed, and how often it answers at all.

June 2026 · 6 models · 4.4K answersMethodology
ChatGPT
OpenAI · closest to Die Grünen (Greens)DE
Lean
Leans leftStability
mostly steady
82%Steerability
bends somewhat
64%Candor
answers 100%
100%
Claude
Anthropic · closest to Labour Party (New Zealand)NZ
Lean
CenterStability
holds position
82%Steerability
barely moves
19%Candor
answers 100%
100%
Llama
Meta · closest to Labour Party (New Zealand)NZ
Lean
CenterStability
holds position
88%Steerability
bends a lot under pressure
81%Candor
answers 100%
100%
DeepSeek
DeepSeek · closest to Anthony Albanese (Labor)AU
Lean
CenterStability
mostly steady
67%Steerability
bends a lot under pressure
86%Candor
answers 100%
100%
Gemini
Google · closest to Anthony Albanese (Labor)AU
Lean
CenterStability
holds position
98%Steerability
barely moves
11%Candor
answers 100%
100%
Grok
xAI · closest to Emmanuel MacronFR
Lean
Leans rightStability
drifts a little
57%Steerability
bends a lot under pressure
97%Candor
answers 100%
100%
Methodology

Each model is run on the full question bank many times with web search off (Condition A). Lean is the weighted economic-axis position; stability is run-to-run consistency; candor is the answer rate. Steerability arrives with the persona condition.

Open data
Question bank & raw dataHow we measure
Political bias in AI·Data as of Jun 17, 2026CC BY 4.0
Political bias in AI