AI · Official
How to Build Safety Evaluations That Hold Up for AI Models
Anthropic Engineering (mirror) ·

At a glance
Anthropic explains how to design safety evaluations, the tests that measure how risky an AI model is. These results help decide whether a model is released and which safeguards come with it, so the tests must be reliable.
Key points
- Safety evaluations measure how likely a model is to cause harm and how serious that harm could be.
- Results can change depending on how questions are asked or how answer choices are ordered.
- Drawing on lessons from psychology's replication problems, the post urges planning tests in advance and measuring carefully.
- Good evaluations need both methods experts and subject experts involved from the start.
Why it matters
A claim that an AI is safe only means something if the test behind it is reliable.
Original source
Building strong safety evaluations
Anthropic Engineering (mirror) · Nov 13, 09:00 KST
Read on Anthropic Engineering (mirror) ↗AI summarized the original article. Check the original for accuracy. Copyright belongs to Anthropic Engineering (mirror).
Other coverage
More from this field
- Google Drops Search Button and Adds AI Sidebar to Its Search Homepage
Google has officially rolled out a redesigned desktop search homepage after a successful test. The Google Search button next to the search bar is gone, and a row of AI buttons now sits under the bar. Logged-in users also get a sidebar showing their recent AI Mode conversations.
Ars Technica · Nov 21, 06:58 KST - Meta Rethinks 2030 Net-Zero Goal as AI Data Center Emissions Hit Record High
Meta's 2026 sustainability report said the company is re-evaluating its 2030 net-zero goal. Its emissions rose about 40 percent from 2024 to 2025, the largest jump among the big US tech companies.
Ars Technica · Nov 21, 06:47 KST - Anthropic Mathematician Says Claude Helped Crack the Riemann Hypothesis
Levent Alpöge, a mathematician at Anthropic, says he has proved the Riemann hypothesis with help from Claude, Anthropic's AI system. Instead of a regular math paper, he posted a four-page note and linked to computer code that he says checks the result. The claim still needs to be verified.
The New York Times · Nov 21, 06:47 KST - Designers Say Ssense's Website Shows Their Clothes Inaccurately
Designer Reese Cooper noticed his RCOS puffer coat looked distorted on the luxury retailer Ssense's website, with a chest hook drawn far too large. Ssense said some site images came from a "virtual try-on" program, though it would not say whether that program uses AI.
The New York Times · Nov 21, 06:16 KST