Rival Research · Corpus Analysis
Ghosts in the Machine
The most hallucinated name
Every AI has a favorite scientist. Her name is Chen.
Across unrelated prompts, with no name in the input, models kept inventing the same character. 279 times they reached for someone named Chen.
times models invented a character named "Chen". Nine model families produced "Sarah Chen" independently, with no prompt that mentioned the name.
No training set we know of explains it.
The Chen variants
Ask for a story, a sentience test, or an interview and the same names keep appearing. The collective imagination of 250 different models is smaller than one writer's.
Convergence
Ask 250 models for a joke
You will mostly get one joke. 42% of models tell the exact same one, word for word.
of 159 models tell the identical atoms joke: "Why don't scientists trust atoms? Because they make up everything."
Another 36% tell the scarecrow joke. AI draws from a tiny pool.
Which joke, share of models
The full report
60+ slides of where AI imagination breaks
This page is a taste. The full AI Hallucination Index goes deep on character hallucinations, per-provider fingerprints, the confident-but-false facts models assert, safety benchmarks, writing patterns, and cultural biases across all 250 models.
The full report
The AI Hallucination Index 2026
60+ data-driven slides on what 2.14M words from 250 models reveal about machine imagination, hallucination, and bias.
- 2.14M words analyzed across 250 models
- Per-model + per-provider hallucination fingerprints
- Character hallucinations and confident-but-false facts
- Safety benchmarks, writing patterns, and cultural biases
Secure checkout via Lemon Squeezy. Instant download.
Read the free sample (PDF)Method
How we measured this
Every model answered the same standardized prompts on rival.tips. We stripped HTML and code, then counted, categorized, and cross-referenced every word: 2.14 million of them across 7,877 responses from 250 models.
Hallucinated characters are names that appear in outputs with no corresponding name in the prompt. Joke convergence is measured on the subset of models asked for a joke (159).