心理学方法揭示AI安全测试的重大缺陷

Researchers at the UK AI Security Institute used psychometric methods to show that popular safety benchmarks for language models don't measure one consistent trait.