Anthropic shares 10 recommendations for designing and reporting AI safety evals
AnthropicAI · x · 2026-11-14
Anthropic published an Engineering Blog post, "Building strong safety evaluations," laying out 10 recommendations for designing, running, and reporting safety evals.
Key points:
- A safety eval measures a model's propensity to cause or contribute to harm—and whether a response makes harm more or less likely—not specific capabilities. It asks: given a specific input, did the model do the right thing?
- It covers both actively harmful requests (e.g., self-harm instructions) and seemingly harmless but high-risk ones (e.g., drug dosing).
- Evals are easy to game, often don't measure what they claim, and are under-documented; results shift with question format and even answer-option ordering.
- The post draws parallels to psychology's replication crisis and endorses reforms like preregistration.
These evals drive deployment decisions, shipped safeguards, and the focus of future safety training and research.
More from Models
- Dev gripes that reasoning models are painfully slow, asks for one-shot era back — rickasaurus · 2026-11-14
- evilsocket hails DeepSeek v4.1-flash-1017 with ARTEX as mind-blowing — evilsocket · 2026-11-14
- Leaked ExploitGym chart hints at GPT-6.2 Astra scoring 2x higher, launch rumored next week — VraserX · 2026-11-14
- NVIDIA unveils GR00T 2, its most advanced open robotics model with cross-body skill transfer — nvidia · 2026-11-14
- Mnemonic Labs (YC F26) launches to train a foundation model for quant trading — ycombinator · 2026-11-14
- New respiratory-virus agent benchmark tops out at 30.2% pass rate — kenbwork · 2026-11-14