Anthropic's TASTE benchmark: best model judges safety proposals at 60%, below human researchers
maksym_andr · x · 2026-11-14
Anthropic's Alignment Science team introduced TASTE (The AI Safety Taste Evaluation), a benchmark measuring how well models judge pairs of AI safety research proposals against experienced researchers' preferences. Two design choices were key to reaching 77% estimated human agreement across 92 pairs: a discussion stage where researchers talk through disagreements before revising scores, and filtering for self-reported strong-confidence labels. The best model, Fable 5, scores just 60%—below human researchers—on this hard-to-verify "research taste" capability that matters for automating safety research. A related METR chart (Benchmark 14) shows near-zero progress across all models on a similar dimension.
Related event: Anthropic's TASTE Benchmark Finds Models Poor at Judging AI Safety Research(3 posts)→
More from Safety
- AI Evaluator Forum Adds Six New Member Organizations — dfrsrchtwts · 2026-11-14
- Open-source bastyn statically scans MCP and coding-agent configs for risks — claudio-rootedlogic · 2026-11-14
- Senate Appropriations advances FY27 bills aimed at securing US AI leadership — peterwildeford · 2026-11-14
- London AI x Crypto scene: cryptographic proofs emerge as key tools for AI verification and governance — DavideCrapis · 2026-11-14
- Researcher predicts cybersecurity-fueled crackdown on open-source models and regulated compute — max_paperclips · 2026-11-14
- FBI FLASH warns threat actors weaponize open-source agentic LLM platform ARTEX — Kyrannio · 2026-11-14