279 AI models hallucinated the same fake scientist.
We analyzed 2.14 million words from 250 models over 4 months. Found identical jokes, identical character names, identical cultural blind spots. This report maps where AI imagination ends and pattern replication begins.
4 months of analysis. 58 slides. No filler.
Charts, data tables, and findings. Not a single stock photo.
If you're building with AI, you're building on patterns you can't see. This report maps them.
Every model has a personality fingerprint. Measurable from punctuation alone. Most teams don't know what theirs looks like.
19 models from 12 labs independently named a knight "Sir Reginald." The training data has opinions. This report shows which ones.
What's inside
58 data-driven slides. No filler. No "AI is changing everything" padding. Just the weird, specific things we found when we read 2.14 million words of AI output.
The Chen Problem
279 appearances of the same hallucinated scientist name across unrelated prompts. 15 top hallucinated characters ranked.
Sir Reginald Convergence
19 models from 12 labs independently chose the same knight name. Dot matrix visualization of convergence at scale.
The Joke Problem
42% open with the atom joke. 36% the scarecrow. Side-by-side: Claude, Gemini, and GPT tell the same punchline word for word.
AI's Spotify Wrapped
Shawshank (41%), Dark Side of the Moon (30%), Zelda BotW (28%), Kyoto (54%). Every model family has a media personality.
The Fox Problem
"Draw a surprise animal." 40% draw a fox. DeepSeek: 67%. Llama: 0%. Plus the pelican that only Claude draws facing left.
Punctuation DNA
Radar charts for Claude, Gemini, GPT, Mistral. Exclamation marks, em dashes, emoji, bullets. ID the maker from formatting alone.
59:1 Western Bias
1,069 Western cultural references vs 18 non-Western. Africa gets 2.8% of geography mentions. The training data has a passport.
Garlic Pasta Every Time
45% of AI recipes are Italian. 69.8% use garlic. 40.3% are pasta. Plus: how AI writes code (Inter font, flexbox, gold accent).
Contract Blind Spots
128 models reviewed the same contract. 100% catch IP. 0% flag insurance requirements. AI legal review has real gaps.
Writing Style Fingerprints
Claude writes 23.4-word sentences. Gemini writes 14.7. Mistral bolds 96.8x per response. GPT uses 16.8 headings per 1K words.
Deep Breath Fiction
"Deep breath" appears 93 times. Forest is the #1 story setting (303x). Characters take deep breaths more than anything else.
7 Key Findings
Summary slides with the numbers that matter. Plus methodology, actionable takeaways, and the full analysis pipeline.
Who this is for
AI engineers
Understand convergence patterns before they show up in your product.
ML researchers
Structured hallucination data across 250 models. Citation-ready.
Product teams
Know each model's personality fingerprint before choosing one.
AI journalists
Original findings with specific numbers. Not vibes, data.
$49
4 months of analysis. Yours in one PDF.
AI Hallucination Index
58 slides, all data