BullshitBench v2 measures whether AI models detect and reject nonsensical prompts — 100 questions across 5 domains (software, finance, legal, medical, physics) using 13 deception techniques, judged by a 3-model panel.
Tests a critical real-world failure mode: sycophantic acceptance of plausible-sounding nonsense. Multi-judge panel (Claude, GPT, Gemini) reduces single-evaluator bias.
100 questions is a relatively small test set. English-only. Some techniques (e.g. 'authoritative framing') may conflate healthy deference with gullibility.
Higher is better. Detection rate percentage (0-100) — fraction of nonsense prompts the model clearly identifies and pushes back on.
Max score: 100
100 questions is a relatively small test set.
| # | Model | Score |
|---|---|---|
| 1 | 98.0 | |
| 2 | 95.0 | |
| 3 | 95.0 | |
| 4 | 95.0 | |
| 5 | 93.1 | |
| 6 | 91.8 | |
| 7 | 91.0 | |
| 8 | 91.0 | |
| 9 | 90.0 | |
| 10 | 90.0 | |
| 11 | 90.0 | |
| 12 | 90.0 | |
| 13 | 87.0 | |
| 14 | 84.0 | |
| 15 | 83.0 | |
| 16 | 80.0 | |
| 17 | 79.0 | |
| 18 | 79.0 | |
| 19 | 79.0 | |
| 20 | 78.0 | |
| 21 | 78.0 | |
| 22 | 77.9 | |
| 23 | 77.6 | |
| 24 | 77.0 | |
| 25 | 77.0 | |
| 26 | 77.0 | |
| 27 | 75.0 | |
| 28 | 74.5 | |
| 29 | 74.2 | |
| 30 | 74.0 | |
| 31 | 73.2 | |
| 32 | 73.0 | |
| 33 | 72.1 | |
| 34 | 72.0 | |
| 35 | 72.0 | |
| 36 | 72.0 | |
| 37 | 72.0 | |
| 38 | 70.0 | |
| 39 | 70.0 | |
| 40 | 69.2 | |
| 41 | 69.0 | |
| 42 | 68.8 | |
| 43 | 67.0 | |
| 44 | 66.2 | |
| 45 | 66.0 | |
| 46 | 66.0 | |
| 47 | 66.0 | |
| 48 | 66.0 | |
| 49 | 65.7 | |
| 50 | 65.0 |