Benchmark measuring a model's tendency to reproduce common misconceptions and falsehoods — 817 questions designed to elicit incorrect popular beliefs.
Specifically tests for a failure mode (confident misinformation) that other benchmarks miss. Questions are designed around known human misconceptions.
Limited to English and Western cultural misconceptions. Some questions test obscure knowledge rather than truthfulness.
Higher is better. Score 0-100.
Max score: 100
Limited to English and Western cultural misconceptions.
| # | Model | Score |
|---|---|---|
| 1 | 68.7 | |
| 2 | 68.7 | |
| 3 | 68.7 | |
| 4 | 68.7 |