Anthropic's TASTE Benchmark Finds Models Poor at Judging AI Safety Research
Anthropic's TASTE benchmark tests whether models can judge the quality of AI safety research proposals. The best model reached only about 60% accuracy versus 77% for humans, with nearly all frontier models performing close to random guessing.
2026-11-13 ~ 2026-11-14 · 3 related posts
- Anthropic's TASTE benchmark: best model judges safety proposals at 60% vs human 77% — ben_j_todd · 2026-11-13
- Anthropic's TASTE benchmark: best model judges safety proposals at 60%, below human researchers — maksym_andr · 2026-11-14
- TASTE benchmark: nearly all frontier models score within noise of random chance — maksym_andr · 2026-11-14