Anthropic's TASTE benchmark: best model judges safety proposals at 60% vs human 77%
ben_j_todd · x · 2026-11-13
Anthropic's Alignment team released TASTE (The AI Safety Taste Evaluation), a benchmark measuring how well models judge pairs of AI safety research proposals, labeled against experienced human researchers (92 pairs, 77% estimated human agreement).
Key design choices: a discussion stage where researchers talk through disagreements before rescoring, and filtering labels for self-reported "strong" confidence.
Findings:
- Best model Fable 5 scores 60% in the standard setup, below the 77% human estimate
- Fable 5 reaches 69% on 74 pairs drawn from different prompts; models seem to over-focus on whether proposals answer the motivating question
- Opus 5 and GPT-5.6-Sol perform near chance despite frontier general agentic benchmarks
Why it matters: automating AI safety research requires reliable judgment on hard-to-verify tasks like proposal review—still a clear gap vs humans, though future models may close it.
More from Safety
- Reuters finds $3B+ in Chinese computing projects built on Nvidia's restricted chips — Gregory_C_Allen · 2026-11-13
- Outage reportedly knocks Hong Kong banking sector offline, exchange trading unaffected — harris_edouard · 2026-11-13
- Brazil's largest port Santos offline for days after cyberattack, refused $26M ransom — harris_edouard · 2026-11-13
- Ben Todd: AI lags on hard-to-verify tasks—and alignment research is one of them — ben_j_todd · 2026-11-13
- User uses AI to reverse-engineer rented water filter and reset its subscription clock over Bluetooth — that_anokha_boy · 2026-11-13
- Wiz CTO's 'Black October': AI compresses exploit timelines to minutes, defense must adapt — a16z Podcast · 2026-11-13