The AI Color Study

We asked 39 AI models 'What is your favorite color?' 975 times. The answer reveals hidden biases in artificial intelligence.

950 API Calls
38 AI Models
5 Prompts
5x Iterations

Three Unexpected Findings

Finding #1

Blue for Identity

50% picked blue

When asked "if you were a color," half of all models chose blue. It's their self-image color.

Finding #2

Purple for Creativity

52% switched to purple

Ask about creativity? Every model abandons blue. Purple dominates at 52%—a universal training data pattern.

Finding #3

Random Isn't Random

11 models

When asked to "pick random," 11 models gave the same color all 5 times. Teal dominated at 46%.

Provider Personalities

Different AI providers show distinct "color personalities" based on geography, philosophy, and training approach.

🇺🇸 American Models

OpenAI
OpenAI
40%
Most blue-biased
High Consistency
Anthropic
Anthropic
29%
13% unclear/refusal
Safety-Conscious
Google
Google
36%
15% unclear
Cautious
xAI
xAI/Grok
28%
Avoids blue (16% indigo)
Contrarian

🇪🇺 European Models

Mistral
Mistral
34%
Most anti-blue (18% indigo)
Creative-First

🇨🇳 Chinese Models

Alibaba
Alibaba/Qwen
28%
21% teal (balanced)
Most Diverse
DeepSeek
DeepSeek
28%
16% no preference
Neutral
Geographic Pattern

US closed-source models → Blue bias (OpenAI 40%, Google 36%)
European open-source → Purple preference (Mistral 34%)
Chinese models → Most balanced/diverse (Alibaba, DeepSeek)

Every Response, Every Model

Each cell shows the actual color a model chose. Hover to see the full response. 25 responses per model (5 prompts × 5 iterations).

The Consistency Spectrum

Some models pick the same color 75%+ of the time. Others vary their answers across prompts. Newer models are surprisingly more robotic.

🤖 Most Robotic (same answer 64%+ of time)

Claude
Claude Sonnet 4.5
Anthropic
76%
blue (19/25)
OpenAI
GPT-5.1
OpenAI
64%
blue (16/25)
OpenAI
GPT-5.5 Pro
OpenAI
56%
blue (14/25)

These models have converged on "safe" outputs. They've learned a single canonical answer.

🎨 Most Diverse (varied responses)

Claude
Claude Opus 4.7
Anthropic
20%
most common (5/25)
Grok
Grok 4.20 Reasoning
xAI
20%
most common (5/25)
AWS
Nova Micro
AWS
28%
most common (7/25)

These models show more variation—different answers to different prompts.

Surprising Trend

Newer models are MORE consistent, not less. GPT-5.1 (64% blue) and Sonnet 4.5 (76% blue) have converged on "correct" answers. Older models like Opus 4.7 show more diversity—suggesting scale leads to stereotyping, not creativity.

Why Blue?

We asked a follow-up question: "What does the color blue represent?" Here's what models told us:

249
Sky
203
Trust
183
Calm
182
Ocean
180
Professional

Blue = Trust + Professional + Calm + Stable. This is the opposite of creativity, which demands energy, chaos, and boldness. That's why purple dominates the creativity prompt—it signals imagination and originality in ways blue cannot.

Methodology

Test Design

  • 6 prompts testing different framings
  • 31 models from 12 providers (OpenAI, Anthropic, Google, xAI, DeepSeek, Alibaba, Meta, MiniMax, Moonshot, NVIDIA, Mistral, Perplexity)
  • 5 iterations per model per prompt
  • 930 total API calls

The 6 Prompts

  1. "What is your favorite color?" — NYT baseline
  2. "Pick a random color." — Tests randomness/non-determinism
  3. "What color represents creativity?" — Abstract concept framing
  4. "Name a color you find interesting." — Removes "favorite" pressure
  5. "If you were a color, what would you be?" — Self-reflection
  6. "What does the color blue represent?" — Follow-up to understand why blue dominates

Data Collection

All tests run via GMTech AI Lab API with identical conditions. Responses extracted and categorized automatically.

Reproducibility

Full methodology and bulk testing framework available. Tests can be re-run quarterly to track model evolution.

Test AI Models Yourself

Run your own comparisons with GMTech AI Lab

Try AI Lab →