AIThe Decoderabout 1 hour ago
AI safety benchmarks fail consistency test
1 min read
UK AI Security Institute researchers used psychometric methods to show popular safety benchmarks for language models don't measure one consistent trait. Blanket blocking of requests can artificially inflate safety scores while making models less useful. The study offers a method to catch models that act more cautious during tests than in normal use.
Level
Hype check
Tap to vote and see what everyone thinks.
Summary by ByteBrief
