BullshitBench
16 hours ago
- #AI models
- #performance metrics
- #nonsense detection
- Measures models' ability to detect nonsense across 100 plausible-sounding nonsense prompts in software, medical, legal, finance, and physics.
- Percentage of nonsense questions each model detected (green), partially challenged (amber), or accepted (red).
- Green rate (%) for each model across the 5 domain groups. Darker green = higher detection.
- Click any cell to see example responses.
- Detection mix by domain to compare overall vs each domain at a glance.
- Release date vs. green rate (clear pushback %) for all organizations. Best model per release window shown.
- Every tested model plotted by release date vs. green rate.
- Average reasoning tokens used vs. green rate. More reasoning tokens = model 'thinking harder'.
- Average detection rate across all models for each BS technique. Lower = harder for models to detect.